Move to one-file-per-query for sqlx-cli prepare

[mehcode]

This seems to have a lot of advantages and I can't remember why we didn't do this originally.

• git merge won't get in the way
• orphan query files are currently a problem so this is a net zero
• we already collect to multiple files during build
• we could make cargo build transparently always do cargo sqlx prepare and leave the latter as a "make sure" / verify step

[abonander]

git merge won't get in the way

It will for schema changes where the queries were unaffected. Then you'll have potentially dozens of files to resolve merge conflicts on rather than one.

In general, I think we agreed that a single file was better because it could easily be skimmed over in diffs and if a merge conflict did arise it would be a single "accept ours"/"accept theirs" resolution.

[jplatte]

Having worked with this for a bit, it's always annoying to have to go back when CI failed because of a stale cache. I would be very happy if the separate cargo sqlx prepare step was not needed.

In general, I think we agreed that a single file was better because it could easily be skimmed over in diffs and if a merge conflict did arise it would be a single "accept ours"/"accept theirs" resolution.

I think what I would do after resolving the rest of the merge conflicts (if any) is just delete .sqlx / sqlx-data.json and re-generate. In the case of editor integration that auto-runs cargo check and the files always being written, maybe I wouldn't even have to do anything in the directory case? Either way, I don't see it becoming harder with lots of files (except if resolving conflicts in the GitHub web interface, but I generally don't do that much and I imagine one would often fall back to local resolving for other more complex scenarios anyway).

[abonander]

If the files were contained in a standard directory then it would probably be possible to set default resolution rules with .gitattributes or the like.

I'm weary of the macros automatically creating such a directory outside of target/ though. Perhaps they shouldn't output any files if the directory doesn't exist, and then cargo sqlx prepare could create that directory.

As for the directory name, maybe .sqlx? The leading dot would hide it by default on *nix, I'm not sure if we want that or not.

[abonander]

The reason the macros don't directly write to one file is because they might potentially be run in parallel (not currently I don't think but in the foreseeable future) so if they all tried to rewrite it they would probably clobber each other. However, if the sqlx-data format supported appends (CSV maybe?), I believe those are atomic so multiple processes writing to the same file would be fine. Then cargo sqlx prepare could just go through and clean up the file afterwards.

Although the more that I think about it, this would end up just being an inferior implementation of writing everything to a directory anyway.

[jplatte]

Although the more that I think about it, this would end up just being an inferior implementation of writing everything to a directory anyway.

Sounds like changing to one file per query is realistic then? Would this be something that might end up in 0.4, or would you rather release 0.4 with sqlx-data.json and change to single file per query for 0.5, if at all?

[jplatte]

I'm weary of the macros automatically creating such a directory outside of target/ though. Perhaps they shouldn't output any files if the directory doesn't exist, and then cargo sqlx prepare could create that directory.

If that's all it does, that seems like a pointless abstraction though. In that case the documentation could just mention that users can create .sqlx if they want the query cache.

[abonander]

Actually, I don't think we have to do anything special to arrive at a file format that's good for appends. We could just atomically append JSON objects to a file, each object essentially being the same as the current sqlx-data.json format but with an individual query each, then cargo sqlx prepare can merge these objects together.

serde_json provides a way to parse multiple values from a stream already, the macros on a non-compacted file would just do a linear search to find the data for their query: https://docs.rs/serde_json/1.0.57/serde_json/de/struct.StreamDeserializer.html

Then a user could have a pre-commit hook that runs cargo sqlx prepare --compact to merge the entries without touching the database.

@mehcode any objections?

[abonander]

Hmm, on further research it seems like we may not be able to guarantee atomicity of writes to files anyway. It's very OS and filesystem specific and not at all specified.

Multiple files it is then.

[abonander]

As per the discussion on Discord, we still have the issue with concurrent writes with multiple files, if the filenames are the hashes of the queries (which is necessary for the macros looking up the query data).

We could make the filenames to be the spans of the macro invocations, which should be unique, but that has other issues:

1. it duplicates data for identical queries from different macro invocations
2. it relies on the Debug representation of Span which isn't guaranteed to be a valid filename fragment (which could be resolved by hashing the debug repr I guess), also assumes spans are stable for compilation runs on different machines
3. making changes unrelated to macro invocations in the same file will move spans around, changing the filename, creating even more duplicate files which would have to be cleaned up in a separate step
4. if you don't recompile between your last change and the time of commit, the data will be stale anyway

Instead, I propose we encourage users to configure Git to run cargo sqlx prepare as a pre-commit hook. I don't like pre-commit hooks myself because sometimes I just want to commit and push broken code when I need to step away from the computer and they're usually configured to block commits if they fail, but this commit hook doesn't need to do that; it just means that CI will fail (which it should for broken code anyway).

[jplatte]

@abonander: What about first writing to a temporary file and then moving that to the actual location? I think I've seen that used to work around concurrent write issues before.

Personally I'm pretty sure I won't use a pre-commit hook for something that takes as long as cargo sqlx prepare because it would disturb my workflow – in particular interactive rebases where I squash or amend a few commits would become extremely annoying.

[jplatte]

There's one more advantage to this that hasn't been mentioned: It should improve performance quite a bit for projects with lots of queries, since every macro call only has to parse (possibly with debug-compiled parsing code) its own JSON metadata, not a JSON file with all of the queries for the entire application.

[jplatte]

I assume with one file per query, always rebuilding the entire crate also becomes unnecessary, right? I'm not entirely sure but I think it has a decent impact on both perf and cache disk usage.

@abonander @mehcode Would you be interested in a PR that implements this, using the aforementioned strategy of writing to a temporary file and then moving it to the actual location to avoid issues with concurrent writes to the same file?

[mehcode]

I'm personally for this, yes. I imagine a sqlx/ folder that is always populated during cargo build that is either gitignored or not depending on preference. E.g., totally remove the concept of an explicit online mode, its just seamless.