Full Width 
Shortcuts 
Sign Up 
Log In 
Articles 
Reading List More from Jaz's Blog
I’ve recently started building a POC of a Redis RESP3 Wire Compatible Key/Value Database built on FoundationDB with @calabro.io and though it’s rather early, it’s already spawned a fun distributed systems problem that I thought would be interesting to share. Previously I’ve written about how I implemented a Graph DB via Roaring Bitmaps, representing relations as a bidirectional pair of sets. To support such use-cases in this new database, we’d like to represent sets of keys such that you can perform boolean operations on them (intersection, union, difference) relatively quickly even for very large sets (with millions of members). Supporting Larger Keys In the original Graph DB, we were representing user DID strings as uint32 UIDs to allow us to store millions of edge lists in very little space (e.g. the set of users who follow bsky.app) while being able to perform boolean operations between lists quickly (using Roaring Bitmaps’ parallel boolean operators). Since we were graphing follows, blocks, and other such User-to-User relationships, there was a practical maximum for the total number of user IDs in the low billions. We’ve continued exploring objects and relationships we’d like to represent as a Graph, and have realized that if we wanted to store e.g. the URIs of all posts a user has liked so we can intersect it with other users’ likes, we’re going to need a bigger keyspace! There are well over 15 Billion records in the AT Proto Ecosystem, each with a unique AT URI! Now our desired keyspace is much larger than can be represented by uint32 values and so we need to expand to uint64. Easy enough, let’s use the uint64 flavor of Roaring Bitmaps and simply intern URIs and User DIDs as uint64s, problem solved, right? Not quite… Interning Many Things at Once The AT Proto Firehose has hit historic peak traffic of over 1,500 evt/sec. We want to design a system that will handle many times more scale than we’ve ever seen in reality. This means designing for 10x or 100x would require us to be able to intern 15k to 150k new URIs per second into uint64 integers. Sounds easy enough, what’s the holdup? Well, in FoundationDB we’re able to use Transactions to do things like atomically increment a sequence safely when many other threads may be trying to do the same thing. This is simple enough to do in Go, we can just toss together a little helper function to acquire a new UID for our string: func (s *server) allocateNewUID(span trace.Span, tx fdb.Transaction) (uint64, error) { var newUID uint64 val, err := tx.Get(fdb.Key("last_uid")).Get() if err != nil { return 0, return fmt.Errorf("failed to get last UID: %w", err) } if len(val) == 0 { newUID = 1 // start from 1 } else { lastUID, err := strconv.ParseUint(string(val), 10, 64) if err != nil { return 0, return fmt.Errorf("failed to parse last UID: %w", err) } newUID = lastUID + 1 } tx.Set(fdb.Key("last_uid"), []byte(strconv.FormatUint(newUID, 10))) return newUID, nil } This function gets called from a fdb.Transaction which gets assigned a Transaction ID, then stages its changes, then tries to commit them. In FoundationDB, if your transaction is reading or modifying data written to by a different Transaction that finishes while you’re in-progress, your Transaction is thrown out and must be retried. For our UID assignment use-case, this is pretty problematic. We want to assign hundreds of thousands of new UIDs per second but if they’re all modifying the same key, concurrent transactions will constantly run into contention on the same data and will be forced to retry over and over again. This problem gets worse the more concurrent transactions you have trying to read from or write to the same key. Even if we stick to sequential access, if it takes ~5-10ms to assign a UID, we can only assign ~100-200 UIDs per second, nowhere near the throughput we need to support. How can we get past this problem and allow us to give strings unique uint64 UIDs in a high throughput and highly concurrent manner? Attempt #1: xxHash My first attempt to solve this problem was to try something that required no coordination and hash the string keys into uint64s using xxHash. xxHash is a non-cryptographic hash algorithm that supports incredibly high throughput (dozens of GB/sec) and can produce 64 bit unsigned integer hashes of strings trivially. Implementing this would look something like: Hash the incoming string key Lookup the uint64 UID to see if we’ve already assigned it to a string Reject the transaction if there’s a collision and give up Store the key in the UID map and the UID in the key map Use the UID for anything else we need While the uint64 keyspace is plenty large for our needs assuming we distribute evenly among the whole space, using a hashing algorithm with no coordination means there’s room for collisions and thus we’d need some additional logic (potentially by bucketing the keys somehow). Consulting the Birthday Problem we can see that a keyspace with 64 bit hashes has a >50% chance of containing a single collision when we have only ~5 billion keys in the set! That’s barely more keys than we can cram into a uint32 and definitely won’t suffice for the number of keys we expect to be storing! So, xxHash, while nice and coordination-free is probably not going to be the solution we need. What else can we do? Attempt #2: Billions of Sequences Incrementing one sequence is clearly not an option because we can only increment a single sequence ~100-200 times per second, but what if we instead had more than one sequence? Roaring Bitmaps managed to make highly efficient bitmap representations by breaking up a uint32 keyspace into a uint16-wide set of uint16-wide keyspaces. Can we do something similar here? Here’s an idea, what if we had just over 4 billion difference sequences and just picked one at random when we needed to assign a UID? Since we’re constructing our UIDs as a uint64, we can split the full UID into a pair of uint32s where the most-significant-bits are used to identify the sequence ID and the least-significant-bits are used to identify the value assigned to the UID within the sequence. So in our implementation, we get ~4.3 Billion sequence IDs that each have ~4.3 Billion incrementing values. As an example, if we were to randomly select Sequence ID 37 and then we increment that sequence to the value 5, we’d assemble the ID as 37<<32 + 5 which looks like 158,913,789,952 + 5 -> 158,913,789,957. Looking at the next Seuqence ID, we’d see 38 which, when left shifted by 32 gives us 163,208,757,248. You can see there’s a gap of ~4.3 billion values between the first UID assigned by each Sequence ID. Assuming we can increment a single sequence ~100 times per second with contention, we’re able to mint 430 Billion new UIDs per second without locking up (assuming the cluster can keep up). Storing ~4.3 billion sequences may be a bit expensive, but thankfully this strategy can scale up and down by picking a larger or smaller prefix size. If we only wanted to store say, ~16k sequences, we can pick a 14 bit prefix instead of a 32 bit prefix and then use a 50 bit sequence number. That spreads the load across 2^14 sequence IDs and significantly reduces storage requirements for Sequences. What does this look like in code? Well, it’s honestly not very complex! const uidSequencePrefix = "uid_sequence/" func (s *server) allocateNewUID(tx fdb.Transaction) (uint64, error) { // sequenceNum is the random uint32 sequence we are using for this allocation var sequenceNum uint32 var sequenceKey string // assignedUID is the uint32 within the sequence we will assign var assignedUID uint32 // Try up to 5 times to find a sequence that is not exhausted for range 5 { // Pick a random uint32 as the sequence we will be using for this UID sequenceNum = rand.Uint32() sequenceKey = fmt.Sprintf("%s%d", uidSequencePrefix, sequenceNum) val, err := tx.Get(fdb.Key(sequenceKey)).Get() if err != nil { return 0, fmt.Errorf("failed to get last UID: %w", err) } if len(val) == 0 { assignedUID = 1 // Start each sequence at 1 } else { lastUID, err := strconv.ParseUint(string(val), 10, 32) if err != nil { return 0, fmt.Errorf("failed to parse last UID: %w", err) } // If we have exhausted this sequence, pick a new random sequence if lastUID >= 0xFFFFFFFF { continue } assignedUID = uint32(lastUID) + 1 } } // If we failed to find a sequence after 5 tries, return an error if assignedUID == 0 { return 0, fmt.Errorf("failed to allocate new UID after 5 attempts") } // Assemble the 64-bit UID from the sequence ID and assigned UID newUID := (uint64(sequenceNum) << 32) | uint64(assignedUID) // Store the assigned UID back to the sequence key for the next allocation tx.Set(fdb.Key(sequenceKey), []byte(strconv.FormatUint(uint64(assignedUID), 10))) // Return the full 64-bit UID return newUID, nil } And there we go! We can now intern billions of strings per second with little to no contention in a distributed system while completely avoiding collisions and making full use of our keyspace! Conclusion Often times when designing distributed systems, patterns and strategies you see in seemingly unrelated libraries can inspire an elegant solution to the problem at hand. In the case of distributed, high-throughput string interning, horizontal scaling can be achieved by breaking up one large keyspace that requires strict coordination into billions of smaller keyspaces that can be randomly load-balanced across. Both patterns used in this technique are present elsewhere: Breaking up a large keyspace into a bunch of smaller keyspaces is present in Roaring Bitmaps (among other systems) Letting randomness and large numbers spread out resource contention is present in many load balancing systems This is one of my favorite parts of growing as an engineer: the more systems and strategies you familiarize yourself with, the more material you have to draw from when designing something new. Personal News A bit of personal news for y’all if you made it this far. Today is my last day as a member of the Bluesky team! The past 2+ years building out Bluesky’s Infrastructure and Platform team and scaling Bluesky from 100,000 -> 40,000,000 users have been the most intense and rewarding years of my life. I don’t have the words to express how much I’ve valued my time on the team and how much I care for the people I’ve worked with in what feels like a decade of real time. I’ve got some new adventures ahead and am excited to be embarking on a new journey within the next month (still building large-scale infrastructure, don’t worry). I plan to continue being involved in the AT Proto Community and to contribute to some cool projects other folks on the Bluesky team are working on from the FOSS space (like KVDB). To the team, I wish you all the best and will dearly miss getting to work with you all every day, but nothing lasts forever and I will always cherish the time I got to spend building an incredible platform with incredible people. If you’re interested in joining a world-class team doing important work, check out Bluesky’s open job listings here. There should be a new role opening up for a seasoned Go Engineer on the Platform team soon!
Often when designing systems, we aim for perfection in things like consistency of data, availability, latency, and more. The hardest part of system design is that it’s difficult (if not impossible) to design systems that have perfect consistency, perfect availability, incredibly low latency, and incredibly high throughput, all at the same time. Instead, when we approach system design, it’s best to treat each of these properties as points on different axes that we balance to find the “right fit” for the application we’re supporting. I recently made some major tradeoffs in the design of Bluesky’s Following Feed/Timeline to improve the performance of writes at the cost of consistency in a way that doesn’t negatively affect users but reduced P99s by over 96%. Timeline Fanout When you make a post on Bluesky, your post is indexed by our systems and persisted to a database where we can fetch it to hydrate and serve in API responses. Additionally, a reference to your post is “fanned out” to your followers so they can see it in their Timelines. This process involves looking up all of your followers, then inserting a new row into each of their Timeline tables in reverse chronological order with a reference to your post. When a user loads their Timeline, we fetch a page of post references and then hydrate the posts/actors concurrently to quickly build an API response and let them see the latest content from people they follow. The Timelines table is sharded by user. This means each user gets their own Timeline partition, randomly distributed among shards of our horizontally scalable database (ScyllaDB), replicated across multiple shards for high availability. Timelines are regularly trimmed when written to, keeping them near a target length and dropping older post references to conserve space. Hot Shards in Your Area Bluesky currently has around 32 Million Users and our Timelines database is broken into hundreds of shards. To support millions of partitions on such a small number of shards, each user’s Timeline partition is colocated with tens of thousands of other users’ Timelines. Under normal circumstances with all users behaving well, this doesn’t present a problem as the work of an individual Timeline is small enough that a shard can handle the work of tens of thousands of them without being heavily taxed. Unfortunately, with a large number of users, some of them will do abnormal things like… well… following hundreds of thousands of other users. Generally, this can be dealt with via policy and moderation to prevent abusive users from causing outsized load on systems, but these processes take time and can be imperfect. When a user follows hundreds of thousands of others, their Timeline becomes hyperactive with writes and trimming occurring at massively elevated rates. This load slows down the individual operations to the user’s Timeline, which is fine for the bad behaving user, but causes problems to the tens of thousands of other users sharing a shard with them. We typically call this situation a “Hot Shard”: where some resident of a shard has “hot” data that is being written to or read from at much higher rates than others. Since the data on the shard is only replicated a few times, we can’t effectively leverage the horizontal scale of our database to process all this additional work. Instead, the “Hot Shard” ends up spending so much time doing work for a single partition that operations to the colocated partitions slow down as well. Stacking Latencies Returning to our Fanout process, let’s consider the case of Fanout for a user followed by 2,000,000 other users. Under normal circumstances, writing to a single Timeline takes an average of ~600 microseconds. If we sequentially write to the Timelines of our user’s followers, we’ll be sitting around for 20 minutes at best to Fanout this post. If instead we concurrently Fanout to 1,000 Timelines at once, we can complete this Fanout job in ~1.2 seconds. That sounds great, except it oversimplifies an important property of systems: tail latencies. The average latency of a write is ~600 microseconds, but some writes take much less time and some take much more. In fact, the P99 latency of writes to the Timelines cluster can be as high as 15 milliseconds! What does this mean for our Fanout? Well, if we concurrently write to 1,000 Timelines at once, statistically we’ll see 10 writes as slow as or slower than 15 milliseconds. In the case of timelines, each “page” of followers is 10,000 users large and each “page” must be fanned out before we fetch the next page. This means that our slowest writes will hold up the fetching and Fanout of the next page. How does this affect our expected Fanout time? Each “page” will have ~100 writes as slow as or slower than the P99 latency. If we get unlucky, they could all stack up on a single routine and end up slowing down a single page of Fanout to 1.5 seconds. In the worst case, for our 2,000,000 Follower celebrity, their post Fanout could end up taking as long as 5 minutes! That’s not even considering P99.9 and P99.99 latencies which could end up being >1 second, which could leave us waiting tens of minutes for our Fanout job. Now imagine how bad this would be for a user with 20,000,000+ Followers! So, how do we fix the problem? By embracing imperfection, of course! Lossy Timelines Imagine a user who follows hundreds of thousands of others. Their Timeline is being written to hundreds of times a second, moving so fast it would be humanly impossible to keep up with the entirety of their Timeline even if it was their full-time job. For a given user, there’s a threshold beyond which it is unreasonable for them to be able to keep up with their Timeline. Beyond this point, they likely consume content through various other feeds and do not primarily use their Following Feed. Additionally, beyond this point, it is reasonable for us to not necessarily have a perfect chronology of everything posted by the many thousands of users they follow, but provide enough content that the Timeline always has something new. Note in this case I’m using the term “reasonable” to loosely convey that as a social media service, there must be a limit to the amount of work we are expected to do for a single user. What if we introduce a mechanism to reduce the correctness of a Timeline such that there is a limit to the amount of work a single Timeline can place on a DB shard. We can assert a reasonable limit for the number of follows a user should have to have a healthy and active Timeline, then increase the “lossiness” of their Timeline the further past that limit they go. A loss_factor can be defined as min(reasonable_limit/num_follows, 1) and can be used to probabilistically drop writes to a Timeline to prevent hot shards. Just before writing a page in Fanout, we can generate a random float between 0 and 1, then compare it to the loss_factor of each user in the page. If the user’s loss_factor is smaller than the generated float, we filter the user out of the page and don’t write to their Timeline. Now, users all have the same number of “follows worth” of Fanout. For example with a reasonable_limit of 2,000, a user who follows 4,000 others will have a loss_factor of 0.5 meaning half the writes to their Timeline will get dropped. For a user following 8,000 others, their loss factor of 0.25 will drop 75% of writes to their Timeline. Thus, each user has a effective ceiling on the amount of Fanout work done for their Timeline. By specifying the limits of reasonable user behavior and embracing imperfection for users who go beyond it, we can continue to provide service that meets the expectations of users without sacrificing scalability of the system. Aside on Caching We write to Timelines at a rate of more than one million times a second during the busy parts of the day. Looking up the number of follows of a given user before fanning out to them would require more than one million additional reads per second to our primary database cluster. This additional load would not be well received by our database and the additional cost wouldn’t be worth the payoff for faster Timeline Fanout. Instead, we implemented an approach that caches high-follow accounts in a Redis sorted set, then each instance of our Fanout service loads an updated version of the set into memory every 30 seconds. This allows us to perform lookups of follow counts for high-follow accounts millions of times per second per Fanount service instance. By caching values which don’t need to be perfect to function correctly in this case, we can once again embrace imperfection in the system to improve performance and scalability without compromising the function of the service. Results We implemented Lossy Timelines a few weeks ago on our production systems and saw a dramatic reduction in hot shards on the Timelines database clusters. In fact, there now appear to be no hot shards in the cluster at all, and the P99 of a page of Fanout work has been reduced by over 90%. Additionally, with the reduction in write P99s, the P99 duration for a full post Fanout has been reduced by over 96%. Jobs that used to take 5-10 minutes for large accounts now take <10 seconds. Knowing where it’s okay to be imperfect lets you trade consistency for other desirable aspects of your systems and scale ever higher. There are plenty of other places for improvement in our Timelines architecture, but this step was a big one towards improving throughput and scalability of Bluesky’s Timelines. If you’re interested in these sorts of problems and would like to help us build the core data services that power Bluesky, check out this job listing. If you’re interested in other open positions at Bluesky, you can find them here.
Bluesky recently saw a massive spike in activity in response to Brazil’s ban of Twitter. As a result, the AT Proto event firehose provided by Bluesky’s Relay at bsky.network has increased in volume by a huge amount. The average event rate during this surge increased by ~1,300%. Before this new surge in activity, the firehose would produce around 24 GB/day of traffic. After the surge, this volume jumped to over 232 GB/day! Keeping up with the full, verified firehose quickly became less practical on cheap cloud infrastructure with metered bandwidth. To help reduce the burden of operating bots, feed generators, labelers, and other non-verifying AT Proto services, I built Jetstream as an alternative, lightweight, filterable JSON firehose for AT Proto. How the Firehose Works The AT Proto firehose is a mechanism used to keep verified, fully synced copies of the repos of all users. Since repos are represented as Merkle Search Trees, each firehose event contains an update to the user’s MST which includes all the changed blocks (nodes in the path from the root to the modified leaf). The root of this path is signed by the repo owner, and a consumer can keep their copy of the repo’s MST up-to-date by applying the diff in the event. For a more in-depth explanation of how Merkle Trees are constructed, check out this explainer. Practically, this means that for every small JSON record added to a repo, we also send along some number of MST blocks (which are content-addressed hashes and thus very information-dense) that are mostly useful for consumers attempting to keep a fully synced, verified copy of the repo. You can think of this as the difference between cloning a git repo v.s. just grabbing the latest version of the files without the .git folder. In this case, the firehose effectively streams the diffs for the repository with commits, signatures, and metadata, which is inherently heavier than a point-in-time checkout of the repo. Because firehose events with repo updates are signed by the repo owner, they allow a consumer to process events from any operator without having to trust the messenger. This is the “Authenticated” part of the Authenticated Transfer (AT) Protocol and is crucial to the correct functioning of the network. That being said, of the hundreds of consumers of Bluesky’s production Relay, >90% of them are building feeds, bots, and other tools that don’t keep full copies of the entire network and don’t verify MST operations at all. For these consumers, all they actually process is the JSON records created, updated, and deleted in each event. If consumers already trust the provider to do validation on their end, they could get by with a much more lightweight data stream. How Jetstream Works Jetstream is a streaming service that consumes an AT Proto com.atproto.sync.subscribeRepos stream and converts it into lightweight, friendly JSON. If you want to try it out yourself, you can connect to my public Jetstream instance and view all posts on Bluesky in realtime: $ websocat "wss://jetstream2.us-east.bsky.network/subscribe?wantedCollections=app.bsky.feed.post" Note: the above instance is operated by Bluesky PBC and is free to use, more instances are listed in the official repo Readme Jetstream converts the CBOR-encoded MST blocks produced by the AT Proto firehose and translates them into JSON objects that are easier to interface with using standard tooling available in programming languages. Since Repo MSTs only contain records in their leaf nodes, this means Jetstream can drop all of the blocks in an event except for those of the leaf nodes, typically leaving only one block per event. In reality, this means that Jetstream’s JSON firehose is nearly 1/10 the size of the full protocol firehose for the same events, but lacks the verifiability and signatures included in the protocol-level firehose. Jetstream events end up looking something like: { "did": "did:plc:eygmaihciaxprqvxpfvl6flk", "time_us": 1725911162329308, "type": "com", "commit": { "rev": "3l3qo2vutsw2b", "type": "c", "collection": "app.bsky.feed.like", "rkey": "3l3qo2vuowo2b", "record": { "$type": "app.bsky.feed.like", "createdAt": "2024-09-09T19:46:02.102Z", "subject": { "cid": "bafyreidc6sydkkbchcyg62v77wbhzvb2mvytlmsychqgwf2xojjtirmzj4", "uri": "at://did:plc:wa7b35aakoll7hugkrjtf3xf/app.bsky.feed.post/3l3pte3p2e325" } }, "cid": "bafyreidwaivazkwu67xztlmuobx35hs2lnfh3kolmgfmucldvhd3sgzcqi" } } Each event lets you know the DID of the repo it applies to, when it was seen by Jetstream (a time-based cursor), and up to one updated repo record as serialized JSON. Check out this 10 second CPU profile of Jetstream serving 200k evt/sec to a local consumer: By dropping the MST and verification overhead by consuming from relay we trust, we’ve reduced the size of a firehose of all events on the network from 232 GB/day to ~41GB/day, but we can do better. Jetstream and zstd I recently read a great engineering blog from Discord about their use of zstd to compress websocket traffic to/from their Gateway service and client applications. Since Jetstream emits marshalled JSON through the websocket for developer-friendliness, I figured it might be a neat idea to see if we could get further bandwidth reduction by employing zstd to compress events we send to consumers. zstd has two basic operating modes, “simple” mode and “streaming” mode. Streaming Compression At first glance, streaming mode seems like it’d be a great fit. We’ve got a websocket connection with a consumer and streaming mode allows the compression to get more efficient over the lifetime of the connection. I went and implemented a streaming compression version of Jetstream where a consumer can request compression when connecting and will get zstd compressed JSON sent as binary messages over the socket instead of plaintext. Unfortunately, this had a massive impact on Jetstream’s server-side CPU utilization. We were effectively compressing every message once per consumer as part of their streaming session. This was not a scalable approach to offering compression on Jetstream. Additionally, Jetstream stores a buffer of the past 24 hours (configurable) of events on disk in PebbleDB to allow consumers to replay events before getting transitioned into live-tailing mode. Jetstream stores serialized JSON in the DB, so playback is just shuffling the bytes into the websocket without having to round-trip the data into a Go struct. When we layer in streaming compression, playback becomes significantly more expensive because we have to compress outgoing events on-the-fly for a consumer that’s catching up. In real numbers, this increased CPU usage of Jetstream by 23% while lowering the throughput of playback from ~200k evt/sec to ~28k evt/sec for a single local consumer. When in streaming mode, we can’t leverage the bytes we compress for one consumer and reuse them for another consumer because zstd’s streaming context window may not be in sync between the two consumers. They haven’t received exactly the same data in the session so the clients on the other end don’t have their state machines in the same state. Since streaming mode’s primary advantage is giving us eventually better efficiency as the encoder learns about the data, what if we just taught the encoder about the data at the start and compress each message statelessly? Dictionary Mode zstd offers a mechanism for initializing an encoder/decoder with pre-optimized settings by providing a dictionary trained on a sample of the data you’ll be encoding/decoding. Using this dictionary, zstd essentially uses it’s smallest encoded representations for the most frequently seen patterns in the sample data. In our case, where we’re compressing serialized JSON with a common event shape and lots of common property names, training a dictionary on a large number of real events should allow us to represent the common elements among messages in the smallest number of bytes. For take two of Jetstream with zstd, let’s to use a single encoder for the whole service that utilizes a custom dictionary trained on 100,000 real events. We can use this encoder to compress every event as we see it, before persisting and emitting it to consumers. Now we end up with two copies of every event, one that’s just serialized JSON, and one that’s statelessly compressed to zstd using our dictionary. Any consumers that want compression can have a copy of the dictionary on their end to initialize a decoder, then when we broadcast the shared compressed event, all consumers can read it without any state or context issues. This requires the consumers and server to have a pre-shared dictionary, which is a major drawback of this implementation but good enough for our purposes. That leaves the problem of event playback for compression-enabled clients. An easy solution here is to just store the compressed events as well! Since we’re only sticking the JSON records into our PebbleDB, the actual size of the 24 hour playback window is <8GB with sstable compression. If we store a copy of the JSON serialized event and a copy of the zstd compressed event, this will, at most, double our storage requirements. Then during playback, if the consumer requests compression, we can just shuffle bytes out of the compressed version of the DB into their socket instead of having to move it through a zstd encoder. Savings Running with a custom dictionary, I was able to get the average Jetstream event down from 482 bytes to just 211 bytes (~0.44 compression ratio). Jetstream allows us to live tail all posts on Bluesky as they’re posted for as little as ~850 MB/day, and we could keep up with all events moving through the firehose during the Brazil Twitter Exodus weekend for 18GB/day (down from 232GB/day). With this scheme, Jetstream is required to compress each event only once before persisting it to disk and emitting it to connected consumers. The CPU impact of these changes is significant in proportion to Jetstream’s incredibly light load but it’s a flat cost we pay once no matter how many consumers we have. (CPU profile from a 30 second pprof sample with 12 consumers live-tailing Jetstream) Additionally, with Jetstream’s shared buffer broadcast architecture, we keep memory allocations incredibly low and the cost per consumer on CPU and RAM is trivial. In the allocation profile below, more than 80% of the allocations are used to consume the full protocol firehose. The total resident memory of Jetstream sits below 16MB, 25% of which is actually consumed by the new zstd dictionary. To bring it all home, here’s a screenshot from the dashboard of my public Jetstream instance serving 12 consumers all with various filters and compression settings, running on a $5/mo OVH VPS. At our new baseline firehose activity, a consumer of the protocol-level firehose would require downloading ~3.16TB/mo to keep up. A Jetstream consumer getting all created, updated, and deleted records without compression enabled would require downloading ~400GB/mo to keep up. A Jetstream consumer that only cares about posts and has zstd compression enabled can get by on as little as ~25.5GB/mo, <99% of the full weight firehose. Feel free to join the conversation about Jetstream and zstd on Bluesky.
More in programming
Download Now View on pouet.net Intro In the late 80s, I spent an embarrassing number of hours hunched over my ZX Spectrum, typing in Sinclair BASIC programs and rudimentary assembler routines copied from magazines, trying to re-create the Amiga crack intros I’d seen round at a friend’s house. Obviously, they never came even slightly close - but now, many decades later, I’ve finally managed it! I’m now the proud author of a hand-rolled retro demoscene intro that runs on a stock 128Kb Spectrum. It features all those late 80s tropes: A smooth scrolltext (with ‘leet custom fonts!), a three-channel AY soundtrack, bouncing VU meters and the essential parallax starfield. It’s currently 10287 bytes of code, or 19552 bytes as a complete tape image - which, given I’ve been working on this since December 13th last year, means my average rate of code-production is a paltry ~ 455 bytes per week! This blog post is a technical overview of how I got there, some cool tricks I picked up, and everything I had to learn (and re-learn) about the Speccy and Z80 assembly along the way. It simultaneously rates as one of my favourite side-projects ever, and also the one that most often nearly triggered a foot put through my monitor in frustration. I’ll reflect on all that at the end, but for now if you just want to see the finished product, the download button is at the top of the article and the included .tap cassette tape image should load into any emulator. There’s also a YouTube capture, links to the production on pouet.net, and you can run it from within a web browser by visiting my TNFS site portal - check option 4 for recent uploads. Be Kind, Rewind But let’s skip back briefly to the late 80s. Coming from the world of an 8-bit, cassette tape loading computer, seeing an Amiga for the first time was like encountering something from outer space. Like everyone who saw one back in the day, I was blown away by the amazing graphics and sound on display, not to mention the library of ground-breaking games. But I found myself even more captivated by the mysterious (and totally badass) “intros” displayed at the beginning of pirated games. Created by the software piracy groups that had cracked the copy protection and distributed the floppy-disk based games, they were mini works of art in their own right - music, graphics and usually some cool-looking effect to show off the group’s prowess. There’s a great collection of these intros (or “cracktros”) archived on YouTube. Watch a few of those videos and you’ll quickly notice the “design” tropes that almost every cracktro shares: a cool, edgy logo at the top of the screen (bonus points if it’s dripping blood or in any way “metal”), some boastful text displayed on top of an effect of some sort - usually a scrolling parallax starfield, or rotating 3d object - and a further scroll-text full of shout outs, greetings to fellow pseudonymous crackers and adverts for illegal BBS systems galore. Needless to say, to a kid in the 80s/early 90s this was possibly the coolest thing ever. I thought this one was the absolute bomb when I first saw it. Sine-scroller, pirate BBS adverts, cool logo with a sword - and of course, a starfield. There’s something for everyone here. The music from this is burned into my brain. I think pretty much everyone of my classmates knew this intro given the popularity of Lemmings in the early 90s. Skid Row as a group are still active to this day and this is probably the quintessential Amiga cracktro of the period. Endless Piracy. Edgy. Another cool logo, sine-wave scroller, starfield and boasting text. The classics never go out of style. Another great combination of all the essential ingredients. Cool logo, even cooler text font, scrolling starfield and a scroll-text running along the bottom. Vintage stuff. Logo. Text. Scroller. But no starfield - otherwise a solid effort from Anthrox here :) The demoscene grew out of these intros. The cracking eventually faded into the background, and coders, “graphicians” and musicians all contributed to feature-length productions that showcased their abilities, competed with other groups and pushed the machine to its limits. In amazing feats of ingenuity and pure “code as an art form”, the scene lives on today, with groups producing breathtaking productions like this and this - bear in mind, both of those are running on a 7Mhz home computer equipped with a floppy disk, 1Mb of memory and a sound chip from 1985. But much as I admire the sheer skill and dedication that takes, it’s the early “single screen” scrolly-text-and-starfield cracktros I’ve always found captivating. They took on a sort of mystical quality for me, hinting at a scene and a culture I could only ever experience from the outside - which is exactly why I spent all those hours in Sinclair BASIC, trying in vain to re-create them, and dreaming of the day I’d eventually have an Amiga of my own. The Jump Back I eventually did trade up to a range of Amiga systems, joined a few “scene” groups myself and threw myself into a variety of activities until exams and the end of my teenage years beckoned. I dabbled with Z80 code back then, and at one point knew enough 68k assembly to bluff my way through - partly from the Amiga’s heyday, partly from a few “embedded systems” modules at University where I had to program a set of traffic lights or lift-controllers. Given there’s not much call for that sort of thing at my $DAYJOB, though, it had all pretty much faded from my memory. I have kept working on a bunch of projects in my spare time, and one of my favourites is my TNFS site - sort of an old-school BBS system or interactive “website” for the Spectrum, which grew out of a fun experiment to see how far I could push Sinclair BASIC. It relies on a modern network card fitted to the Spectrum, which is how I came to be chatting to “Desertkun”, another TNFS-protocol enthusiast. He’s now producing a great new WiFi network card for the Speccy, and was kind enough to send me a few of the prototypes to test out for him. I was massively inspired by his hardware and low-level coding chops, and during one of our video calls where we’d been comparing our development setups, he asked why I didn’t write any Z80 assembly myself. A throw-away comment, but the challenge was set! Now, I couldn’t face re-writing the whole site in asm, but I figured I could produce a self-contained intro to promote it in the style of the early “BBStros”, and finally be one of the cool kids with a scene production of my own :) I mean, how hard could it be ? (ooooh - foreshadowing!) Dev Environment My development environment through this was pretty straightforward, and is the same stack I use for developing my TNFS site. I just used a Makefile-driven build system which ran my code through the Pasmo assembler, my fork of zmakebas to build BASIC files, and the FUSE emulator for testing. For final testing on the real hardware, I uploaded my .tap files to my local TNFS server where I could run them on my +3. A few additional tools I have to call out because they ended up making a big difference - one is the amazing web-based “asm-editor” I discovered while looking up Z80 instructions. Not only does this site have a great reference of all the Z80 instructions, but it also has an editor where you can run example code and see how the memory and/or various registers change state. Take for example, the djnz instruction - you can build it and step through the example loop construct one instruction at a time. Or try my example of an add instruction in the screenshot to see how the carry flag works when a bit gets shifted out of a register. Although it uses a slightly different syntax to Pasmo (so I couldn’t drop my code directly in it), it massively helped when manually stepping through things. And lastly, probably the most significant debugging tools I used on this project - a notepad and pen, and my two dogs (say “Hi!” to Biscuit & Pebble…) I’ll write some more about this in the wrap-up, but this really was a highly unusual project for me, coming from decades of much higher-level languages and constructs. Most of my z80 code I only ever really understood after drawing out the bytes and bitmasks, keeping a table of register flags and drawing chunks of screen memory out by hand. And I found my evening walks with the two dogs provided ideal rubber duck exercises, although I’m not sure what they both made of my protracted mutterings about the hl register pair. Either way, debugging-by-dog is highly recommended! The Starting Point So, let’s get started. I followed a great “Hello, world!” example from futurice.com to validate my build environment. Beyond that, the first thing I wanted to do was load in my logo and background image I’d use as a “stage” for the remainder of the planned intro. I wasn’t really sure what I was going to do at this point, just that I wanted my logo at the top, and a scrolltext along the bottom row of the screen. As my TNFS site header already looked kinda demoscene-y, I made a quick edit putting in place two blue guide-bars inside which I’d put further content. Memory map courtesy of breakintoprogram.co.uk A Spectrum SCREEN$ file (usually displayed as loading screens) is just a dump of screen memory, starting from byte 16384, and running for 6912 bytes. The weird layout of this is very relevant later, but in order to display it all you need to do is load your file into that memory location. After a few hour’s work of reviewing Speccy internals and following various guides, I had a very simple starting point for my demo that served as the basis for everything else that came after. org $8000 screen EQU $4000 ; address of the first byte of screen memory, 16384 in dec scrlen EQU 6912 ; total length of the screen memory inc. attributes ula EQU $FE ; I/O port for the ULA to control the border black EQU 0 ; Spectrum colour number, the exact same ones BASIC uses start: ld a,black ; Set the border colour to black using the ULA out (ula),a ; Copy my logo header picture to screen memory ld hl,header ; HL = location of my screen file in this code ld de,screen ; DE = start of screen memory ld bc,scrlen ; BC = 6912 bytes to copy ldir ; block copy, see https://jnz.dk/z80/ldir.html ; busy loop - rest of intro goes here :P hold: halt ; sleep until the next interrupt (1/50 second) jr hold ; then jump back and sleep again, for ever header: incbin "assets/header.scr" ; my logo header/background end start Not much to look at now, but this simple starting point introduced me to important concepts like interfacing with the Spectrum’s hardware (ULA chip to start with, and AY sound chip later) via the out instruction, how to pull in binary files using Pasmo’s INCBIN directive, using the Z80’s block copy ldir instruction, the use of register pairs for 16-bit addressing, and simple branching to labels. Hello World The next step was to print something on the screen and explore some of the Speccy’s built-in ROM routines, which are somewhat analagous to syscalls on modern operating systems. There are plenty of disassemblies and guides to the Spectrum ROM available on the internet, one of the best commented and linked one is at https://skoolkid.github.io/rom/. Using this as a reference alongside the various “Hello, world!” tutorials showed me how this sort of thing is built up: at EQU 22 ; "at" control character used to position the cursor row EQU 11 ; coordinates where to print the message col EQU 9 prstring EQU $203C ; where the pr_string routine lives in ROM ld de,msg ; DE = address of the first character ld bc,msglen ; BC = how many characters to print call prstring ; print the lot ; busy loop is the same... msg: db at,row,col ; equivalent of "PRINT AT 11,9" in BASIC db "Hello, world!" msglen EQU $-msg ; 16 bytes in all, pasmo shortcut, $=program counter I ended up mostly not using the ROM routines in my code as they dragged in all sorts of things I didn’t need, but it was massively useful to understand how they worked as they introduced me to yet more commonly-used constructs. For example, the PR_STRING routine defined here shows how you can build a loop over each character. If you’re just starting out, I highly recommend going over the 8 lines in that routine with a notepad and pen, working out the contents of the registers and following the loop constructs. I also made a version of it in asm-editor you can try. If you click that link, you’ll see a version of the code loaded into the web editor. Click “Build”, and then step through the code using the “Step” button - although no text will be actually displayed (I skipped over the call to the Spectrum ROM “restart” vector that does this), you’ll see the various states of the registers and for example the BC register counting down as each byte of the string is processed. Adding Some Style Problem is, the Speccy’s built-in font is not exactly dripping with demoscene credibility. I wanted to use some stylish fonts and this turned out to be super easy - a quick search turned up Jim Blimey’s awesome font collection, which included a ready-to-use set of assembler files and some instructions which soon led me to these additions in my code: ; CHARS is the address the ROM's printing routine reads to find the font. ; Picking our own font is simply a matter of over-writing this chars EQU $5C36 ; from Jim's docs - POKE 23607=5C36 in hex ld hl,myfont-256 ; HL = font address - 256 as per docs ld (chars),hl ; from now on the ROM draws with it! ; rest of code is the same myfont: include "textfont.asm" ; "Cartoon" font from jimblimey.com There we go! Now that’s looking a little more elite. Sorry, I mean 31337. Strap in folks, we’re leaving lamer city… So at this point, a month or so in to my dabblings and I’d started to get a reasonable understanding of common idioms and was starting to feel like this was a project that might actually go somewhere. The problem started when I moved on to trying to implement actual demo effects: It became very apparent, very quickly, just how much I didn’t know about how this all worked. Not to mention the fact that at this level, you’re really closely exposed to the physical hardware intricacies, including the design decisions and trade-offs that the hardware designers had to make when working with the extremely limited specs of the early 1980s. Cursed Screen Layout The first such problem - and the cause of probably 90% of my pain - is the Speccy’s extremely weird screen layout. Rather than standard bitmapped display you might expect on more modern hardware, it is instead split into 3 sections, each with a set of interleaved rows (so if you keep adding 1 to a screen address, eventually instead of reaching the next row, you’ll instead end up 8 rows down, and so on until you hit the bottom of the section and then loop back to the row underneath where you started!) followed by a block of “attribute” bytes which determine the colours of each 8x8 block. The attribute blocks are the reason why each 8x8 character cell can only have two colours (a PAPER and INK attribute) which led to the garish “colour-clash” of spectrum games and distinctive style where artists try to keep inside 8x8 blocks as much as possible. But it was the layout of the bitmap portion that is such a mind twister. Download the video You can see this illustrated whenever a game reads in a loading screen from tape. You can see each section built up with the interleaved rows one at a time, and then finally the colour and other attributes (like flashing ). There’s a lot of detail at the fantastic Break Into Program site. In the defense of the designers, they were working with extremely limited hardware and I don’t think the Speccy was ever imagined as a games machine. When you start shuffling bytes around, you realise that this approach is really optimised for displaying text which is neatly arranged into 8x8 character boundaries. For example, using this “scrambled” bit order, if you have a screen address in the hl register pair, you can quickly get to the next pixel line of a given address by simply incrementing the “high” byte. There’s a bunch of other “cheap moves” you can make by simply shifting bits around: Move Instruction Right 8 pixels inc l Left 8 pixels dec l Down 1 pixel line inc h Up 1 pixel line dec h Down 1 character row add 32 to l It was a reasonable choice for a machine that would probably spend a lot of it’s time displaying BASIC program listings - it’s only when you start trying to write demo or sprite routines that it starts to feel really weird. This also means you’ll pretty quickly need a routine to convert X,Y addresses to the relevant memory addresses. Again, L Break Into Program has a useful routine and if you disassemble more or less any spectrum game or demo, you’ll find a version of this procedure somewhere. And yes, this all became horribly relevant later on… The Scroller Download the video Admittedly, not the most interesting demonstration - but this was the first time I had made something move on the Spectrum’s screen with assembly! I watched this with pride more times than I care to admit :) I had a snazzy logo and “Hello World” text, but that doesn’t make for a very interesting intro! The first effect I wanted was the scrolltext - without that, it’s not really an intro. Now, the Amiga had all kinds of hardware sprites, blitter operations, custom chips but the Spectrum had none of that. I was going to have to work out how to do things the hard way - “banging on the metal”, literally by shifting bits around the screen memory by hand and having the CPU do all the work. This is where understanding the Speccy’s weird screen and character cell layout is so important. It turns out there’s a neat trick as long as I stay inside one character row - there’s a RL instruction, which takes a byte and “rotates” it left. It inserts whatever is in the carry flag into the rightmost bit, and whatever gets shifted out ends up in the carry bit again. That means if you string a bunch of them together, you can move a set of pixels across the screen - the moment they “fall off” one character cell, they end up in the carry flag and then the next rl instruction inserts them at the beginning of the next character cell to the left. Repeat 32 times, and you’ve just moved a pixel (or set of pixels) across a line of the screen: linestart EQU $4880 ; leftmost byte of the line 12 rows down lineend EQU linestart+31 ; rightmost byte column ($489F) linelen EQU 32 ; one whole line is 32 bytes (character cells) across dot EQU %00000001 ; the dot "shape", just one pixel lit at the far right end newdot: ld a,dot ; load the dot shape (only rightmost pixel lit) into a ld (lineend),a ; store it in the rightmost byte of the line, pixel is now lit frame: halt ; wait for the next frame, otherwise it whizzes across too fast to see ld hl,lineend ; hl = last (rightmost) byte of the line ld b,linelen ; loop counter, 32 bytes to work through or a ; trick to clear the carry flag, or against itself leaves a register ; unchanged but clears various bits including the carry flag ; see https://jnz.dk/z80/or_r.html shift: rl (hl) ; rotate this byte left (00000001 -> 00000010) dec hl ; point hl at the next byte to the left djnz shift ; loop construct: b=b-1 and then jump back to shift label until b=0 jr c,newdot ; now we've shifted the whole line, so if the carry flag is set now it means ; the dot has just left the screen, so we jump back to the start jr frame ; otherwise it is still somewhere on the screen, so keep shifting! Download the video Hell yeah. It’s almost a proper scrolltext now! Wrapping this in another loop that increments h each time (remember - that gets us moving down a scanline) gets us moving an entire character cell. Here you can see it moving an entire row of my “Hello world” message. There was however a lot more work needed to make a proper scrolltext, with each letter appearing smoothly. I started off by just printing an entire character at a time on the right-hand edge of the screen - it worked, but was pretty ugly as you saw each new character appear at once in the right-hand side of the screen, scrolling 8 pixels over and then the next character flashed up. Instead, I wanted to make each character appear pixel by pixel and that required a bit more bit juggling. The routine I ended up with effectively streams each character in, line by line, pixel by pixel using the font data directly instead of calling the ROM print character routine. It does this by looking up the ASCII code for a character, and then fetching the corresponding font representation into a 8-byte buffer, one byte for each line of data. The font data starts with the space character (character code 32), so a lookup is simply font code base address + (ascii code - 32) * 8 to get to the bits for that letter. Arranged into an 8x8 grid, the bits for the letter H in the font data look like this. For each row of bits, I shift the line left using a neat trick. The byte is loaded into the a register, and then by adding itself to itself - add a,a we shift the value left (in binary, multiplying by 2 moves everything over one bit) so that the pixel I want to draw ends up dropping into the carry flag. This is an idiomatic way of doing it in Z80 code as it’s faster and smaller than the sla instruction. And every cycle quickly adds up when you’re re-drawing 256 bytes every frame! The character buffer then looks like this, with the top line shifted. Now, with the hl register pair pointing at the correct location on the screen, I can then use the rl trick discussed above. This shifts a scanline from the bottom character row of the screen left by one pixel and draws my pixel in the rightmost position by moving it in from the carry flag. Looping 32 times over this moves the entire row by 1 pixel with my new pixel inserted at the right. I then repeat this step again using inc h to move down one scan line until the entire character column has been drawn. At this point, the character buffer looks like this, and all 8 lines of the bottom screen row have all been shifted left by one bit by looping over all 32 character columns. The whole thing then repeats another 7 times until the entire character has been smoothly scrolled onto the screen. So just to re-state this because I think it’s worth the emphasis - just to scroll a message across the bottom of the screen, we’re having to loop over and re-draw 256 pixels every frame! It shows how much faster assembly is than the BASIC interpreter built into the spectrum’s ROM. I found it amazing to see my code finally working and imagining all the bits flying around and registers flapping away. It really makes me appreciate just how much effort went into creating the games I played back then. Music The original Amiga module I converted. A proper 1980s ear-worm! Of course, any good intro needs some music. On the Amiga, the standard was the “tracker” module format, which has a long and involved history of different forks, versions and tweaks. A song is built up from a series of patterns, each made up of rows of instructions. Each instruction usually includes note or pitch information, which sample to play and various parameters such as volume, effects - or in the case of some instructions, which position in the song to jump to. In many ways, it’s almost like a programming language for music. Using ProTracker as an example, a typical note cell is 4 bytes, shown on-screen as something like C-2 01 C20. That breaks down into three fields: the note and octave (C-2), the sample number (01), and the effect and its parameter (C20 - in this case, “set the volume to $20”, which is half of the maximum $40). A close-up of Protracker playing a tune. Note the VUmeters - another classic demo effect I would pinch later on! For an impressive demonstration of what’s possible with a tracker and to see what the note data looks like on-screen, check out the work of h0ffman, and in particular his winning “Professional Tracker” entry from the Revision 2014 competition. He’s an undisputed master of the science (or art!) of tracking, and his “Secrets of Protracker” talk gives a great overview of what’s possible. The Amiga’s sound hardware far eclipsed anything else on the market at the time, and the ZX Spectrum is paltry by comparison. The original 48K models only offered a rudimentary “beeper”, but the later 128K models included a 3-channel AY-3-8912 sound chip. This was very popular in 8-bit micros of the time, and a more-or-less identical variant was also used in the Atari ST. While it only offers basic waveforms and sound processing, it does enable multi-channel “chip music”, and so it became the mainstay of the ZX Spectrum demoscene. It also spawned a variety of SoundTracker clones on the Spectrum itself, although most musicians these days prefer to work in modern tools like Vortex Tracker II, which runs natively under Windows. The Conversion Pex looking back on his demos - this starts at the point where he features my letter and plays a bit of my cover version! For my intro, I wanted to port an Amiga module - and I had the perfect one in mind. One of my favourite Amiga music disks was “Sounds Of Gnome”, released in 1989 by the Swedish duo “Mahoney & Kaktus”. Mahoney also wrote NoiseTracker, one of the first really widely used trackers on the Amiga, and the program that served as my teenage introduction to making computer music of my own. Back in 2018 I made a rock cover of one of the other tunes on that disk, and wrote to Mahoney (real name “Pex”) about it. He ended up showing my “music video” at a demo conference, which I am eternally grateful for! So it seemed only fitting that I mangle another song from that collection with a Spectrum conversion… Sorry, guys! I picked the track “Milkman” which was written by “Glue Master”, an associate of Mahoney’s who had a few of his tracks featured on the compilation. It’s an ideal track for converting: It’s short (only 7 patterns in total), most of those patterns share the same chord sequence, and it looked fairly easy to squeeze the original 4 channels down into the Spectrum’s 3. Plus, it’s an incredibly catchy tune and I knew it well - I’ve been listening to it for 35+ years and often jam along to it on guitar or bass. The conversion started off pretty easily: all I really had to do was copy the patterns note-for-note from the Amiga mod, which I had open in OpenMPT as my Windows tracker. Below is a session with both trackers open on the same pattern - the original Amiga mod on the left, and my 3-channel Spectrum conversion on the right. Note the 11.7 in the 3rd channel of the Spectrum version. This is a “pitch slide down” effect in Vortex Tracker syntax, which drops the pitch by about a quarter-tone. It’s repeated on the next line, so the overall effect is a slide down from the A# that leads in to the next bass note. The original Amiga mod does the same thing using its own syntax - those are the 202s you can see on the same notes, at the bottom of the 4th track. Using just a single “beep” sound for everything was enough to check everything was working, and after a few weeks of hacking on it and re-arranging things in the evenings, I had a basic version which sounded surprisingly good for a first draft: Download the MP3 instead. I have to give a special mention to this forum post which was a massive help getting me started. It’s probably the best introduction to Vortex Tracker I’ve found - many, many thanks to the original author for taking the time to write it all up! The real challenge was making it sound as close to the original as possible. That meant building up a library of better-sounding samples for the different parts, and making extensive use of Vortex Tracker’s “ornaments”. Vortex Tracker comes with a great selection of sounds, and fortunately I’m also a weirdo who likes to listen to chiptunes in the background while I work! So whenever I found something particularly enticing, I filed it away for later pilfering inspiration. A Vortex Tracker sample isn’t really a “sample” in the Amiga sense of a recorded sound - it’s a definition of how a note sounds, one line per frame. Each line can mix in a tone or the white noise generator (useful for snare effects and suchlike), along with various other parameters. Ornaments, on the other hand, give you extra control over what pitch a note plays - and this enables really useful tricks like “arpeggio chords”. For example, an ornament of 0,4 makes the channel switch rapidly (25 times a second) between the note being played and the one 4 semitones above it. It happens so quickly that it sounds much like two notes played at the same time, which means you can build up full-sounding triad chords using only two channels. Take the opening chord in my fully converted version, Cm: It plays C in the first channel and the minor 3rd (D#) in the second, then I apply an ornament to the second channel to bring it up 4 semitones to the 5th (G). That gives all three notes of the chord, while keeping the 3rd channel free for other parts. As well as making much more efficient use of the limited 3 channels, this also gives Spectrum music that characteristic “warbling” sound, which I find really appealing in a retro kind of way! Between tricks like this and samples that mix tones with white noise (to make a combined “bass-snare” sound, for example), I managed to condense the full track down to 3 channels. Many months later, I have to say I’m really pleased with the final result. Playback The original site that was home to Vortex Tracker, documentation and other resources appears to be offline at the moment. I managed to grab a copy of everything I needed from archive.org - you’ll want the PTxTools.7z file for the playback routines discussed here. The playback was simple in theory: As this overview notes, you basically just pull in the player routine, include your module file and call the playback routine once per frame in your main loop. The catch is, none of the source code versions supplied in the archive worked with my Pasmo assembler. Fortunately, there is a pre-compiled binary object shipped alongside the source (VTII10bG in the PTxTools/PT3Play directory). This just has to be loaded in at address $C000 - this cannot be changed, as it has lots of absolute addresses baked in. My code to handle module playback now looks something like this, with the binary module loaded in place: ; Define the start address of the player routine and functions we need player EQU $C000 ; hard-coded org address of binary player plinit EQU player+3 ; init function in docs, module address in hl plplay EQU player+5 ; play and move on one frame ; from the docs - this "analyzes PT3 module header and prepares ; corresponding note and volume tables" ld hl,moduleimg call plinit ; main loop hold: halt call plplay call scroll ; other effects / routines go here... jr hold moduleimg: incbin "assets/music.pt3" Starfield Download the video Well, that didn’t work as hoped. TNFStro - tortoise edition. This brings me to the main event. No self-respecting cracktro is complete without a starfield! My first, naive approach to the classic starfield effect was to scatter random stars across the screen, then move down each scan-line applying the same rl trick I’d used for the scroll-text. I figured I could get parallax by shifting each row a different amount: One pixel for the first row, two for the second, three for the third, then back to one again, all the way down the screen to give the impression of several “fields” of stars moving at different speeds. It turned out there were two major problems with that approach. Firstly, shifting an entire row smeared whatever else was on it along with the stars - which I’d expected, and had filed away as a problem for future me. The bigger issue was speed: by the time I’d finished the first three rows I’d already performed hundreds of shift operations, and it was nowhere near fast enough to fit inside a single frame. Everything crawled across the screen, and the music playback slowed down so much that you could actually hear the arpeggio ornaments warbling away as separate notes. Clearly I was going to have to keep track of the individual stars instead. I found this routine, which was more-or-less what I had in mind and started me off with the idea of a structure to hold the state of each star. It also uses a screen address lookup much like the one from “Break Into Program”. Download the video Dammit. Back with another funeral dirge rendition :( That got things moving, but the co-ordinate lookup and tracking proved expensive to do (it has to run for every star, every frame as well as keep track of the old position for erasing) and I’d been ignoring the bigger question how to display stuff on top of the starfield. My original idea was simply to draw the stars and then paint text over them each frame, which blew the frame budget just as comprehensively as my first attempt. If I wanted something more interesting than an empty starfield, I’d have to keep track of which cells were occupied by text and avoid drawing stars in them in the first place, rather than over-painting them every time. After lots of scribbling in my notepad and visualising rows of bits, I eventually solved both challenges by “cheating” fairly liberally. So now, I don’t generate a random starfield at all. Instead, I built a table of pre-calculated Y and X position pairs: stdata: defb 109, 25, 88,182, 42, 96, 155,226 defb 59,254, 137,169, 160, 84, 125,193 ; ... and so on, for 32 stars ... This saves faffing about with an RNG, and once I have a layout that looks good I know it’ll always look good. I can also “fudge” things and ensure no two stars share the same y co-ordinate which sidesteps a whole class of problems like stars sharing a scanline, and having to deal with them overtaking each other. The “random” layout is generated by a simple Ruby script: #!/usr/bin/env ruby PAIRS = 32 PER_LINE = 4 # pairs per line Y_RANGE = 40..175 # exclude rows taken up by my logo & bars X_RANGE = 0..255 raise "y range too small" if Y_RANGE.size < PAIRS # no duplicate y values ys = Y_RANGE.to_a.sample(PAIRS) pairs = ys.map { |y| [y, rand(X_RANGE)] } pairs.each_slice(PER_LINE) do |row| body = row.map { |y, x| format('%3d,%3d,', y, x) }.join(' ') puts "defb #{body.chomp(',')}" end When the routine initialises, I convert those co-ordinate pairs into a table that tracks the state of each star. So rather than storing a position and translating it to a screen address every frame, each of my star records now looks like this (offsets from the base address): STSCR EQU 0 ; screen address of the byte the star is in, e.g. $4123 STOCC EQU 2 ; a pointer to the "occupancy check" table cell STMASK EQU 4 ; the star's bit pattern, e.g. %00010000 STSTEP EQU 5 ; "speed" - num. of pixels this star travels per frame The key thing is that there are no X/Y co-ordinates at all. For each star I store the screen byte it sits in and the bit mask within that byte, so it’s already in the form the screen writes need. That means no address conversion per frame, just simple bit shifting and rotation. Moving a star is then dead simple. I rotate its mask left with rlca, and when the bit rotates out of bit 7, it lands back in bit 0 with the carry flag set. That’s the signal I need to move the star one character cell to the left, which is one of the “cheap moves” from earlier. Because I’m tracking which cell each star is in, the collision check comes almost for free. That’s what the second pointer is for: STOCC (Star Occupancy) moves in step with STSCR through a table holding one byte per character cell. The “typewriter” routine sets a byte to 1 for every cell containing a text character or the cursor, so checking whether a star is hidden behind text is a single or instruction. If the cell is occupied, I skip drawing the star but carry on rotating and stepping it along until it reaches a cell where it can be displayed: ld l,(ix+STOCC) ld h,(ix+STOCC+1) ld a,(hl) ; A = the occupancy check for this cell or a ; test A against zero jr nz,stmove ; it's occupied, so leave the screen alone Parallax is still handled by moving each star a different number of pixels per frame, but the cost is a lot lower this time as I’m only touching the cells the stars actually occupy, rather than every byte of every row. Download the video That’s looking more like it! The typewriter routine is a simple copy of the font data from memory, and is very cheap as we can use the “cheap moves” to quickly fetch the 8x8 block. I also skip drawing any space character so it leaves the “occupancy cell” unmarked. This means the stars keep drifting through the gaps in the text - you can see here the odd star passing in between the words. I switched to ALL CAPS writing for the final version which helps the effect even more as the font is quite blocky, so you don’t really notice stars skipping the spaces inside a glyph at all. Misc Notes A few odds and ends that didn’t warrant a section of their own… VU Meters Some sort of graphic display bouncing along with the audio playback was another common early demo trope. I did originally consider a real-time graphic equaliser, but that would have been far too complex, probably in terms of spare cycles per frame, and definitely in terms of my limited mental capacity at this point! So I settled for a simple set of bouncing meter bars instead. It’s worth a mention mostly because it gives me an excuse to link to the awesome chibiakumas website! I just love the late 90s web design on display - one can argue that readability and accessibility take a hit, but it’s just so much fun to encounter a site that still looks like this. It’s a clear labour of love and there’s a wealth of content there as well, if your eyeballs are up to it. Their guide on Making Sound with the AY-3-8910 was what showed me how to get the data I needed to draw the meter bars. The examples write a register number to $FFFD to select a register, and I realised that once a register is selected, I could use the in instruction to read its value back out instead of writing a new one. As the page says, R8, R9 and R10 hold the channel amplitudes, so they feed nicely into the height parameter for drawing a bar. Getting the amplitude for channel A is then as simple as: ld a,8 ; Channel A register call ayread ; ... rest of code here ... ayread: ld bc,$FFFD ; Register select port out (c),a in a,(c) ret Twinkle, Twinkle, Little Star This is a simple effect that makes the header logo a little more interesting. I’m not sure where I originally got the idea - I’ve definitely seen something like it on a few late 80s/early 90s demos, though. All I wanted was to pick random star dots in the logo background and animate them. Here’s the header image loaded in ZXPaintbrush with the character cell grid showing. By happy chance, there are four stars which each sit neatly inside a single character cell with enough surrounding space, so they can be used to give an animated “twinkle” effect by drawing a cross out from the centre point and back in again. These dots map to the following co-ordinates, which target the top scanline of each fully-grown star (two pixels up from the centre): twkdata: defb 1,5 ; char row 0, col 5, centre y=3 ink 5 defb 1,20 ; char row 0, col 20, centre y=3 ink 1 bright defb 26,7 ; char row 3, col 7, centre y=4 ink 4 defb 25,22 ; char row 3, col 22, centre y=3 ink 6 bright The “twinkle” routine itself is simple enough: it adds a pixel on either axis to grow the star, waits a few frames, then shrinks it back down again. The interesting bit was how I made the twinkle order pseudo-random. Remember that my starfield was entirely pre-calculated from a data table generated by a Ruby script, which meant I’d never actually needed a random number generator in assembly. From Jonathan Cauldwell’s “How To Write Spectrum Games” e-book, I found another great “hack” which simply walks through the start of the Spectrum’s ROM, picking out random-looking bytes as it goes. As he says, he uses it in most of his games, and it does the job in a really simple way - as long as you stay below the block of structured, repeating data that sits 8Kb into the ROM. twkrand: push hl ld hl,(twkseed) ld a,h and 31 ; keep it within the first 8k of ROM ld h,a ld a,(hl) ; the "random" byte inc hl ld (twkseed),hl pop hl ret Out Of Space About halfway through writing the intro, I hit a fairly predictable problem: I had run out of memory. In an attempt to keep everything self-contained in one block of code, I was pulling all of the assets (the screen background, the player binary, the music file etc.) into the program itself, and then using the ldir block copy instruction to move each one into its correct location. Since that meant effectively storing every block twice, and the player routine had a hard-coded start address of $C000, I quickly ran out of room and my code started crashing into the playback routine. The simplest option was to split things out into separate code blocks, load each one directly to the address it needed to live at, and wrap the lot in a multi-stage BASIC tape loader: 10 CLEAR 32767 20 BORDER 0: PAPER 0: INK 0: CLS 30 LOAD "" CODE 40 LOAD "" CODE 50 LOAD "" CODE 60 RANDOMIZE USR 32768 This loads the main program, then the player routine directly into $C000, and finally the background screen straight into screen memory. It does mean the background starts appearing before the demo has fully loaded, but as I imagine most people will be seeing this running in an emulator or on my TNFS site, it doesn’t really matter. Plus, at this point I was in danger of disappearing down yet more rabbit holes (I still really want to work out how to do multi-colour border effects!) and I’d never have got this thing released. The Zen Of Z80 This whole project is a really weird one to look back at. On the one hand, I’ve fulfilled a childhood ambition and have something I’m very proud of. It’s also pretty unusual these days to end up with a binary where you can say pretty much what every single byte is doing and the story of how it got there. But on the other hand - I now have an answer for why I don’t code in assembly: It’s because I value (what’s left) of my sanity! The thing is, in order to write code like this you have to start thinking like the machine itself. You have to visualise the registers and the status flags flapping away, graphics as lines of bits slotted into strange orders, loops and conditionals as binary logic and so on. Getting this close to the hardware definitely brought about a greater sense of attachment to the hardware of my childhood, and it’s easy to see why assembly coders and sceners get so passionate about it. That said, I’m in no hurry to start re-writing my body of work in assembly! The projects I really enjoy are those that give me an excuse to learn something for the sheer hell of it, and to slow down a little from the ever-increasing velocity and demands of the modern work world. Going back to Z80 code is probably about the furthest I could have taken this, short of going into hardware hacking. I found it at times an intriguing puzzle, often a frustrating exercise that nearly had me throwing in the towel at several stages. This project has certainly taken the longest to go from the “idea” stage of any given feature to any kind of meaningful output. True, I didn’t spend every day of the last year working on it - my actual coding time was probably only a few hours over a couple of evenings a week. But I’ve certainly been thinking about it and pondering it most days (as my dogs will attest). The biggest breakthroughs and “aha!” moments usually didn’t come from behind the keyboard; they came from scribbles in my notebook and thinking about processor states while doing something unrelated. It has however given me massive appreciation and respect for the coders who pull this stuff off and make it look easy. And that goes double for those kids and bedroom coding superstars who managed all of this back in the 80s, and had to work it all out by experiment themselves. I only managed to get this far because of numerous books, online tutorials, GitHub repositories with sample code, YouTube videos… I can’t imagine how much longer this would have all taken me if the only resources I had were my Spectrum, a cassette tape recorder and a notepad. So, as I draw a line under this project - for now! - I raise a glass, and in the words of h0ffman: “Shout out to the hackers, graphicians and protrackers.”
A decade ago, a little bit of history was made. I didn't realize it, but a colleague made a great point, one of those real mind-changing points that seem too obvious to admit same-day. But, the next day, calver.org was born. At the time my team maintained the Python infrastructure for eBay and PayPal, and we were stuck deciding whether we were really ready for a "major" 1.0 release. Semantic Versioning was the only game in town and "major" means "big", right?! Thankfully, a wiser colleague mentioned: Ubuntu and Twisted don't struggle with version number debates. They slap a date on it and keep shipping. In fact, their date-based versions were even better because you always knew where it stood, in terms of updatedness and support. The only problem is that no one really knew about it. Somehow, this problem solving versioning alternative, arguably as old as history itself, had gone nameless for millenia, conspiring to make me feel foolish in an office meeting. Never again! Ten years of adoption Fast forward 10 years, we've seen CalVer adopted by Apple, Nvidia, JetBrains, and countless others. (We have a timeline!) The site may have more inbound links than any other project of mine. Apple made the biggest jump, at WWDC 2025: iOS went from 18 to 26 macOS from 15 to 26 watchOS from 11 to 26 and visionOS from 2 to 26 All landing on one, consistent number like a car's model year. I still remember the texts from the Venn diagram fanbase of my friends who love Apple and reasonable versioning. No such texts from when NVIDIA announced calendar versions across the GPU Operator, RAPIDS, and its monthly NGC containers, but still very cool. Open source, too: Home Assistant, pip, CockroachDB, and yt-dlp all ship on dates, with plenty more on the users page. The conversation even reached the language core; PEP 2026 proposed versioning CPython as 3.YY, and it almost happened, too. And it's never too late, time marches on! Fixing the notation But I don't think I got every detail right from day 1. That's the main motivator for CalVer 26. It's high time to start righting a couple idiosyncratic token design choices, starting with some additions: Meaning Before 26.0 26.0 Full year YYYY YYYY Short year (6, 16) YY YY Zero-padded year (06, 16) 0Y 0Y Short month (1 ... 12) MM M Zero-padded month (01 ... 12) 0M 0M Short week (1 ... 52) WW W Zero-padded week (01 ... 52) 0W 0W Short day (1 ... 31) DD D Zero-padded day (01 ... 31) 0D 0D Seeing double First, the doubled letters. From the first version (16.6), MM and DD meant the unpadded month and day, which reads backwards to anyone who knows date formats (ISO 8601's YYYY-MM-DD, Java, moment.js, day.js), as some community members correctly pointed out. I was ready to flip them, until I checked what people actually use: most projects with a YY.MM.MICRO badge (conda, Twisted, Ansible's tooling) don't pad, and more than a dozen other version management tools (like bumpver and bump-my-version) implement the old meaning. So, it's too late to flip MM's meaning. Instead, 26.0 deprecates it and offers a more explicit and hopefully clearer option: M is the short month, 0M the padded one, and MM is a technically-retired synonym for M. In case you're wondering, the explicit 0M was me being overinspired by Ubuntu's approach, perhaps: 6.06 pads its month but not its year, and YY.0M says exactly that. To pad or not to pad I think it's worth a detour into why padding is even a thing anyways. It's become important now that new ecosystems have emerged that enforced SemVer formatting semantics, and I wanted clear guidance about on the spec site. SemVer forbids leading zeros outright, so Cargo rejects 26.04.0 and Go modules reject v26.04.0. Even Python's packaging spec normalizes leading zeros away, so you can tag 2026.08.19 if you want, but PyPI will still show 2026.8.19. NVIDIA's GPU Operator docs put it this way: "Zero padding is omitted for month to be still compatible with semantic versioning." CalVer was always intended to drop in where SemVer was used. So 26.0 recommends unpadded (YYYY.M.D) as a sane "pure" default for software libraries. But libraries are not the only objects of versioning schemes. The exception is a version that becomes a filename, an image tag, or an object-store key that gets listed lexically. There, padding keeps 26.10 sorted after 26.09, which is why Ubuntu, NixOS, and NVIDIA's own NGC containers pad. More evidence of teams designing their versions. We love to see it. Our FAQ has a longer discussion of the padding issue, as well. Optional segments There was never any rule against them, but 26.0 makes optional trailing segments more explicit with square brackets. Now, yt-dlp's scheme can finally be written down: YYYY.0M.0D[.MICRO]. For the CalVer badges I could find on GitHub, they all stay valid for now. I've got a note on the deprecated spellings and a new copy-paste badge section for new ones. What else is new? It's always a great time to add more citations to the site. A spec changelog; the spec now versions itself: 16.6, 19.7, 26.0. A FAQ: breaking changes, same-day releases, and padding. The users page, rebuilt by category, with past users of note (schemes change; that's fine) and tooling. Case studies: Apple and NVIDIA in, yt-dlp replacing youtube-dl. Much of the thinking behind these changes happened in the GitHub issue tracker over the years, and 19 or so issues close with this release. That's where ideas for CalVer should go, so by all means, open an issue, and we'll get it sorted! In due time, of course. In closing, I can't believe I still love belaboring these numbers so much. Thanks to all (but especially Mark, Glyph, Hugo, issue reporters, translators, and maintainers) for the discussion, and ultimately making the most timely versioning system a timeless classic. See also 2016 announcement Designing a version My Yap on Why CalVer beats Semver
Four sincere attempts at OKRs, four failures, and every time: "you didn't do it right." Maybe the framework doesn't fit you -- so how do you find one that does?
Multiple concurrent writers. Multiple processes. Same SQLite. No modifications.
Brilliant jerks, ZIRP-era managers, and how psychological safety lost the plot. Part 3 of my conversation with Dr. Cat Hicks.
