SQLite's documentation about its durability properties is unclear
agwa.name
agwa.name
> More broadly, I don't think a single definition of 'durable' (as in ACID D) for transactions is particularly useful.
> Much more useful is to ask "what kinds of failures could cause committed transactions to be lost?"
All these articles talking about durability as a singular term should be warned. Writing to disk is not durable under certain circumstances. fsync is not durable under certain circumstances. Two-phase commit is not durable under certain circumstances. Multi-data center commit is not durable against the Death Star too.
1) MicroSD card started producing bad sectors. 2) Power failure during write. 3) Fraudulent MicroSD being used.
Settings section of the database where written so few times that they could be recovered. It was the journaling of events for analysis that where the ones to be lost. Most of the time it was a page or two that was corrupted and the database worked just fine until detailed reporters where created.
Settings where also mirrored to a backup location. Unless the client wanted to pay more for history backup it was only retained locally.
SQLite (with WAL) doesn't do `fsync` on each commit under default settings - https://news.ycombinator.com/item?id=45005071 - Aug 2025 (90 comments)
with a relevant comment by the creator of SQLite here: https://news.ycombinator.com/item?id=45014296
(via https://news.ycombinator.com/item?id=45068594 - thanks int_19h!)
Perhaps what's unclear is when to select which option?
Also, even if I've understood the docs correctly, a number of people in this thread and elsewhere have come to a different interpretation. I think that's much less likely to happen with the PostgreSQL docs.
And I'm sure you can understand why I began to doubt my own interpretation of the docs when SQLite's creator posted a comment saying the exact opposite of what I thought the docs said!
I think it does come down to the definition of durable. The default in SQLite is that the data for a transaction will be written all the way to disk (with an fsync) and the rollback journal will be deleted (but without an fsync).
In this model, there is chance of losing only the last transaction and only if the whole system crashes (like a power failure) after deleting the journal but before your file system makes that durable. If your application crashes, you're still fine.
That's significantly more durable than an asynchronous commit in Postgres, though. In an asynchronous commit, the transaction will complete before any data is written to disk at all - the data is only in memory, and many transactions can queue up in memory during the wal_writer_delay. All of those will definitely be lost if Postgres stops uncleanly for any reason (for example a crash, oomkill, or power failure).
> If you switch to WAL mode, the default behavior is that transactions are durable across application crashes (or SIGKILL or similar) but are not necessarily durable across OS crashes or power failures. Transactions are atomic across OS crashes and power failures. But if you commit a transaction in WAL mode and take a power loss shortly thereafter, the transaction might be rolled back after power is restored.
https://news.ycombinator.com/item?id=45014296
the documentation is in contradiction with this.
From the documentation, it seems like synchronous being FULL does provide durability of the database in DELETE mode, as FULL means it calls fsync after the transaction is completed. I think you may be confusing durability of the journal file with durability of the database. I don't think WAL can ever really have a durable transaction; it is essentially based on leaving a transaction open until it gets "check-pointed" or actually committed to the database file.
In general: WAL means you write the transaction to the WAL, fsync (in sqlite, this depends upon the sync mode], and then return it's done to the application. The transaction is then durable: even if the database crashes, the contents of the WAL will be applied to the database file.
Checkpointing later just lets you throw away that part of the WAL and not have to replay as much to the database file.
If the directory containing the rollback journal is not fsynced after the journal file is deleted, then the journal file might rematerialize after a power failure, causing sqlite to roll back a committed transaction. And fsyncing the directory doesn't seem to happen unless you set synchronous to EXTRA, per the docs cited in the blog post.
I don't follow. How would fsyncing the rollback journal affect the durability of the actual database? Do you actually think that the database would reapply an already committed journal whose ID in the header already indicates that the transaction was committed, when the database is already consistent? I really think you should re-review the definition of durability of a database, especially before saying the creator of SQLite is incorrect about its implementation.
FWIW I don't think it's wrong per se. The article links to a HN comment in which Richard Hipp explains why this is the default behavior, and it does make sense: https://news.ycombinator.com/item?id=45014296. At the same time, clearly, the definition of "durable" here could use some clarification.
Note that the comment by Richard Hipp is justifying why WAL mode is not durable by default. It's a completely reasonable explanation, and would be for DELETE mode too, yet his comment claims that DELETE mode is durable by default, which I can't reconcile with the docs.
Guys. The journal would not be a hot journal though, as the hot journal selection only applies if the database is in a inconsistent state. Otherwise, the database knows from the ID of the journal not to reapply an already applied rollback journal. The process you are talking about ONLY happens when the journal database has been corrupted state, and it has to try and file a file to help recover the database.
I believe they would all be satisfied.
I don't see any mention of checking IDs. Not saying you're wrong - I think the docs could very well be wrong - but could you provide a citation for that behavior?
If you're not going to provide citations for your claims, yet criticize me for "picking out sections" when I provide citations, then continuing this conversation won't be productive.
But it's not "already applied", that's the whole point. The transaction was committed, not rolled back, so the changes in transaction were persisted to disk and the journal was just thrown away. If it magically reappears again, how is SQLite supposed to know that it needs to be discarded again rather than applied to revert the change?
In terminal 1, I created a database and added a table to it:
$ sqlite3 testdb
sqlite> create table test (col int);
$ gdb sqlite3 `pidof sqlite3`
(gdb) b unlink
(gdb) c
sqlite> insert into test values(123);
$ cp testdb-journal testdb-journal.save
(gdb) c
Back in terminal 3, I sent SIGKILL to sqlite3, simulating a power failure:
$ killall -9 sqlite3
$ mv testdb-journal.save testdb-journal
This proves int_19h and I are right - if the journal file comes back, SQLite will apply it and roll back a committed transaction.
I then confirmed with strace that, as the documentation says, the directory is only fsynced after unlink when synchronous=EXTRA. It doesn't happen with synchronous=FULL. So you need synchronous=EXTRA to get durability in DELETE mode.
It's possible I've misunderstood how DELETE mode works. But here's the thing - I shouldn't have to understand how DELETE mode works to know what SQLite setting I need to use to get durability. Unfortunately, the SQLite docs don't clearly say what guarantees each setting provides - instead they talk about about what SQLite does when you choose the setting, leaving the reader to try to figure out if those actions provide durability. And the docs really make it seem like you need synchronous=EXTRA in DELETE mode to get durability, for the reasons explained above.
This is a docs problem; I'm not saying SQLite is buggy.
I think this is the part that is confusing.
The fsyncing of the directory is supposed to be done by the filesystem/OS itself, not the application.
From man fsync,
As well as flushing the file data, fsync() also flushes the metadata information associated with the file (see inode(7)).
Unfortunately it seems like this has traditionally been broken on many systems, requiring workarounds, like SYNCHRONOUS = EXTRA.
The next paragraph in the man page explains this:
> Calling fsync() does not necessarily ensure that the entry in the directory containing the file has also reached disk. For that an explicit fsync() on a file descriptor for the directory is also needed.
https://man7.org/linux/man-pages/man2/fsync.2.html
Edit to add: I don't think there's a single Unix-like OS on which fsync would also fsync the directory, since a file can appear in an arbitrary number of directories, and the kernel doesn't know all the directories in which an open file appears.
This is a moot point anyways, because in DELETE mode, the operation that needs to be durably persisted is the unlinking of the journal file - what would you fsync for that besides the directory itself?
I was more curious so I looked at the code here:
https://sqlite.org/src/file?name=src/pager.c&ci=trunk
and found something similar to what you are asking in this comment before `sqlite3PagerCommitPhaseTwo`:
** When this function is called, the database file has been completely
** updated to reflect the changes made by the current transaction and
** synced to disk. The journal file still exists in the file-system
** though, and if a failure occurs at this point it will eventually
** be used as a hot-journal and the current transaction rolled back.
** This function finalizes the journal file, either by deleting,
** truncating or partially zeroing it, so that it cannot be used
** for hot-journal rollback. Once this is done the transaction is
** irrevocably committed.
Right.
> Assuming fsync works on both the main database and the hot journal, then I don't see a way that it is not durable? Because, it has to write and sync the full hot journal, then write to the main database, then zero out the hot journal, sync that, and only then does it atomically return from the commit? (assuming FULL and DELETE)
DELETE mode doesn't truncate or zero the journal; it merely deletes it from the directory. You need to switch to TRUNCATE or PERSIST for that behavior: https://sqlite.org/pragma.html#pragma_journal_mode
I confirmed all of this by attaching gdb to SQLite and setting a breakpoint on unlink. At the time of unlink the journal is still "hot" and usable for rollback: https://news.ycombinator.com/item?id=45069533
** journalMode==DELETE
** The journal file is closed and deleted using sqlite3OsDelete().
**
** If the pager is running in exclusive mode, this method of finalizing
** the journal file is never used. Instead, if the journalMode is
** DELETE and the pager is in exclusive mode, the method described under
** journalMode==PERSIST is used instead.
I know we have used SQLite quite a bit without getting into this exact scenario - or at least the transaction that would have been rolled back wasn't important enough in our case, I guess when the device powers down milliseconds later to trigger this case we never noticed (or needed this transaction), only that the DB remained consistent.
And it seems like if DELETE is the default and doesn't get noticed enough in practice that it needs to be changed to a different safer default (like to PERSIST or something,) I guess it is better to have the speed gains from reducing that extra write + fsync.
But, now I guess you know enough to submit better documentation upstream for a sliding scale of durability, I definitely agree that the docs could be better about how to get ultimate durability, and how to tune the knobs for `journal_mode` and `synchronous`.
why not? if you use synchronous=FULL, then WAL does provide durable transactions, no?
An application that really wants confidence in a write—to the extent that the underlying device and drivers allow—should use O_DIRECT. Or maybe there is a modern equivalent with io-uring. But that is not easy engineering :)
But I'm not sure I'd necessarily think of O_DIRECT as a way of improving "confidence in a write". It's a way to get a specific behavior.
What OP is referring to is that some drives ignore that signal for performance reasons and there’s nothing SW can do to solve that part.
io_uring in no way changes the rules here.
v3.0 was first released in 2004—over 20 years ago—and the industry has changed dramatically since then.
I can’t help but wish for a “v4.0” release: one that deliberately breaks backward compatibility and outdated defaults, in order to offer a cleaner, more modern foundation.
Note: I'm not asking for new functionality per se. But just a version of SQLite that defaulted to how it should be used, deployed in 2025.
[^1]: https://en.wikipedia.org/wiki/Log-structured_merge-tree
The page you link even mentions scenarios they know about that do happen and that they still assume won't happen. So even sqlite doesn't make anywhere near as strong a claim as you make.
> SQLite assumes that the operating system will buffer writes and that a write request will return before data has actually been stored in the mass storage device. SQLite further assumes that write operations will be reordered by the operating system. For this reason, SQLite does a "flush" or "fsync" operation at key points. SQLite assumes that the flush or fsync will not return until all pending write operations for the file that is being flushed have completed. We are told that the flush and fsync primitives are broken on some versions of Windows and Linux. This is unfortunate. It opens SQLite up to the possibility of database corruption following a power loss in the middle of a commit. However, there is nothing that SQLite can do to test for or remedy the situation. SQLite assumes that the operating system that it is running on works as advertised. If that is not quite the case, well then hopefully you will not lose power too often.
And? If you write to a disk and later this disk is missing, you don't have durability. SQLite cannot automatically help you to commit your writes to a satellite for durability against species ending event on Earth, and hence its "durability" has limits exactly as spelled out by them.
I don't think that is often done now.
Were this a one off, you would have a point. It isn't, however. My experience over many years has been that you can't ever be certain about what is actually going on, based on the documentation alone, and that you wind up in Reddit and Stack Overflow and a plethora of blog posts attempting to figure it out. With LLMs, we have only more sources of contradictory and chronically obsolescent input.
There is an actual problem here. However I can see that, based on the contributions from the SQLite downmod mafia, this talk isn't welcome, so I'm off to some other thing. Have a nice weekend, I suppose.
Oh, come on. There's no open or secret attempt at censoring talk about sqlite on HN. (The story is #11 on the frontpage the minute the comment was made.)