Transactions are not an intrinsic part of a storage system. Any storage system can be made transactional: Redis, S3, the filesystem, etc. Delta Lake and Orleans demonstrated techniques to make S3 (or cloud storage in general) transactional. Epoxy demonstrated techniques to make Redis (and any other system) transactional. And of course there's always good old Two-Phase Commit. If you don't want to read those papers, I wrote about a simplified implementation of Delta Lake and also wrote about a simplified MVCC implementation over a generic key-value storage layer. It is both the beauty and the burden of transactions that they are not intrinsic to a storage system. Postgres and MySQL and SQLite have transactions. But you don't need to use them. It isn't possible to require you to use transactions. Many developers, myself a few years ago included, do not know why you should use them. (Hint: read Designing Data Intensive Applications.) And you can take it even further by ignoring the transaction layer of an existing transactional database and implement your own transaction layer as Convex has done (the Epoxy paper above also does this). It isn't entirely clear that you have a lot to lose by implementing your own transaction layer since the indexes you'd want on the version field of a value would only be as expensive or slow as any other secondary index in a transactional database. Though why you'd do this isn't entirely clear (I will like to read about this from Convex some time). It's useful to see transaction protocols as another tool in your system design tool chest when you care about consistency, atomicity, and isolation. Especially as you build systems that span data systems. Maybe, as Ben Hindman hinted at the last NYC Systems, even proprietary APIs will eventually provide something like two-phase commit so physical systems outside our control can become transactional too. Transactions are a protocol short new post pic.twitter.com/nTj5LZUpUr — Phil Eaton (@eatonphil) April 20, 2025

There are a few interesting scenarios to keep in mind when writing applications (not just databases!) that interact with read and writes files, particularly in transactional contexts where you actually care about the integrity of the data and when you are editing data in place (versus copy-on-write for example). If I don't say otherwise I'm talking about behavior on Linux. The research version of this blog post is Parity Lost and Parity Regained and Characteristics, Impact, and Tolerance of Partial Disk Failures. These two papers also go into the frequency of some of the issues discussed here. These behaviors actually happen in real life! Thank you to Alex Miller and George Xanthakis for reviewing a draft of this post. Terminology Some of these terms are reused in different contexts, and sometimes they are reused because they effectively mean the same thing in a certain configuration. But I'll try to be explicit to avoid confusion. Sector The smallest amount of data that can be read and written atomically by hardware. It used to be 512 bytes, but on modern disks it is often 4KiB. There doesn't seem to be any safe assumption you can make about sector size, despite file system defaults (see below). You must check your disks to know. Block (filesystem/kernel view) Typically set to the sector size since only this block size is atomic. The default in ext4 is 4KiB. Page (kernel view) A disk block that is in memory. Any reads/writes less than the size of a block will read the entire block into kernel memory even if less than that amount is sent back to userland. Page (database/application view) The smallest amount of data the system (database, application, etc.) chooses to act on, when it's read or written or held in memory. The page size is some multiple of the filesystem/kernel block size (including the multiple being 1). SQLite's default page size is 4KiB. MySQL's default page size is 16KiB. Postgres's default page size is 8KiB. Things that go wrong The data didn't reach disk By default, file writes succeed when the data is copied into kernel memory (buffered IO). The man page for write(2) says: A successful return from write() does not make any guarantee that data has been committed to disk. On some filesystems, including NFS, it does not even guarantee that space has successfully been reserved for the data. In this case, some errors might be delayed until a future write(), fsync(2), or even close(2). The only way to be sure is to call fsync(2) after you are done writing all your data. If you don't call fsync on Linux the data isn't necessarily durably on disk, and if the system crashes or restarts before the disk writes the data to non-volatile storage, you may lose data. With O_DIRECT, file writes succeed when the data is copied to at least the disk cache. Alternatively you could open the file with O_DIRECT|O_SYNC (or O_DIRECT|O_DSYNC) and forgo fsync calls. fsync on macOS is a no-op. If you're confused, read Userland Disk I/O. Postgres, SQLite, MongoDB, MySQL fsync data before considering a transaction successful by default. RocksDB does not. The data was fsynced but fsync failed fsync isn't guaranteed to succeed. And when it fails you can't tell which write failed. It may not even be a failure of a write to a file that your process opened: Ideally, the kernel would report errors only on file descriptions on which writes were done that subsequently failed to be written back. The generic pagecache infrastructure does not track the file descriptions that have dirtied each individual page however, so determining which file descriptors should get back an error is not possible. Instead, the generic writeback error tracking infrastructure in the kernel settles for reporting errors to fsync on all file descriptions that were open at the time that the error occurred. In a situation with multiple writers, all of them will get back an error on a subsequent fsync, even if all of the writes done through that particular file descriptor succeeded (or even if there were no writes on that file descriptor at all). Don't be 2018-era Postgres. The only way to have known which exact write failed would be to open the file with O_DIRECT|O_SYNC (or O_DIRECT|O_DSYNC), though this is not the only way to handle fsync failures. The data was corrupted If you don't checksum your data on write and check the checksum on read (as well as periodic scrubbing a la ZFS) you will never be aware if and when the data gets corrupted and you will have to restore (who knows how far back in time) from backups if and when you notice. ZFS, MongoDB (WiredTiger), MySQL (InnoDB), and RocksDB checksum data by default. Postgres and SQLite do not (though databases created from Postgres 18+ will). You should probably turn on checksums on any system that supports it, regardless of the default. The data was partially written Only when the page size you write = block size of your filesystem = sector size of your disk is a write guaranteed to be atomic. If you need to write multiple sectors of data atomically there is the risk that some sectors are written and then the system crashes or restarts. This is called torn writes or torn pages. Postgres, SQLite, and MySQL (InnoDB) handle torn writes. Torn writes are by definition not relevant to immutable storage systems like RocksDB (and other LSM Tree or Copy-on-Write systems like MongoDB (WiredTiger)) unless writes (that update metadata) span sectors. If your file system duplicates all writes like MySQL (InnoDB) does (like you can with data=journal in ext4) you may also not have to worry about torn writes. On the other hand, this amplifies writes 2x. The data didn't reach disk, part 2 Sometimes fsync succeeds but the data isn't actually on disk because the disk is lying. These are called lost writes or phantom writes. You can be resilient to phantom writes by always reading back what you wrote (expensive) or versioning what you wrote. Databases and file systems generally do not seem to handle this situation. The data was written to the wrong place, read from the wrong place If you aren't including where data is supposed to be on disk as part of the checksum or page itself, you risk being unaware that you wrote data to the wrong place or that you read from the wrong place. This is called misdirected writes/reads. Databases and file systems generally do not seem to handle this situation. Further reading In increasing levels of paranoia (laudatory) follow ZFS, Andrea and Remzi Arpaci-Dusseau, and TigerBeetle.

This is an external post of mine. Click here if you are not redirected.

This is an external post of mine. Click here if you are not redirected.