Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

feat(server): library refresh go brrr #14456

Draft
wants to merge 9 commits into
base: main
Choose a base branch
from
Draft

Conversation

etnoy
Copy link
Contributor

@etnoy etnoy commented Dec 2, 2024

This PR significantly improves library scanning performance. Wherever suitable, we are doing jobs in batches, and many looped database interactions are replaced with SQL queries.

Benchmark 1
A library scan with 22k items where nothing has changed since the last scan used to take 1m 22s, now it's below 10 seconds, an improvement of 87 percent!

Benchmark 2
A clean library import with 19k items takes 1m40s in main and 7 seconds in this PR.
NOTE: this benchmark is only the library service scan and does not include the metadata extraction. Also, some fs calls have been migrated from the library service to the metadata service, although this should only have a minor impact on overall scan performance

Benchmark 3
Importing a library with >5M assets.

  • Time to 1M imported (without metadata extraction): 6m50s

Highlights:

  • File paths crawled on disk are compared in sql to discard already-imported files
  • Modified files are scanned in batches, then a single db call updates all of them
  • Missing files are identified in batches, then a single db call marks all of them as offline
  • Import paths and exclusion patterns are matched against library assets in a single sql query

Bonus:

  • Greatly improved log messages related to library scans
  • More e2e tests for handling when offline files go back online, leading to one major bug fixed

TODO

  • More e2e tests for the in-db exclusion pattern matching
  • Test the library watcher

TODO PRs
We need to branch out the following PRs before finishing this one

  • Don't delete offline assets from disk when emptying trash
  • Allow a user to remove offline assets from trash
  • Allow undefined dates in the asset table

@etnoy etnoy force-pushed the feat/inline-offline-check branch from 80aa615 to 8ecde3b Compare December 2, 2024 21:46
Copy link
Contributor

@mertalev mertalev left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nice start! I think there are still a lot of untapped potential improvements here.

server/src/services/library.service.ts Outdated Show resolved Hide resolved
server/src/services/library.service.ts Outdated Show resolved Hide resolved
@mertalev
Copy link
Contributor

mertalev commented Dec 4, 2024

The update to fileCreatedAt, fileModifiedAt and originalFileName is unnecessary and can be handled in metadata extraction since this will be queued anyway. This makes the batched update for isOffline and deletedAt simpler since there'll be no values that are unique to each asset.

@etnoy
Copy link
Contributor Author

etnoy commented Dec 8, 2024

Thanks for your comments @mertalev ! I'll first attempt to do the import path and exclusion pattern checks in SQL and then move to your suggestions

@etnoy etnoy force-pushed the feat/inline-offline-check branch 2 times, most recently from d394654 to 8b2a48c Compare December 9, 2024 21:34
@etnoy etnoy force-pushed the feat/inline-offline-check branch 3 times, most recently from 6d69307 to c26f6aa Compare December 10, 2024 16:41
@etnoy etnoy force-pushed the feat/inline-offline-check branch from c26f6aa to a3be620 Compare December 10, 2024 20:39
@etnoy etnoy changed the title feat(server): run all offline checks in a single job feat(server): library refresh go brrr Dec 10, 2024
.where({ isOffline: false })
.andWhere(
new Brackets((qb) => {
qb.where('originalPath NOT SIMILAR TO :paths', {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Use LIKE instead of SIMILAR TO.

The exclusions and import paths are also specific to a particular library, right? So you need to specify the library in the query.

Also, can you generate SQL for this and confirm with EXPLAIN ANALYZE that it uses an index?

.update()
.set({
isOffline: true,
deletedAt: new Date(),
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The status also needs to be set. This is why I don't really like the status field. The same info is stored in multiple places so it's so easy for it to go out of sync like this.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm trying to remember how the status field works. We use it to distinguish between user-deleted and offlined items, right? What shuld it be set to here?

@etnoy etnoy force-pushed the feat/inline-offline-check branch 5 times, most recently from 775b817 to 69b273d Compare December 12, 2024 20:59
@etnoy
Copy link
Contributor Author

etnoy commented Dec 12, 2024

The update to fileCreatedAt, fileModifiedAt and originalFileName is unnecessary and can be handled in metadata extraction since this will be queued anyway. This makes the batched update for isOffline and deletedAt simpler since there'll be no values that are unique to each asset.

Never thought of that, I've implemented your suggestion. I'm also considering changing the initial import code to ignore file mtime, this allows us to not do any file system calls except for the crawl. Metadata extraction will have to do the heavy lifting instead

@mertalev
Copy link
Contributor

mertalev commented Dec 12, 2024

The update to fileCreatedAt, fileModifiedAt and originalFileName is unnecessary and can be handled in metadata extraction since this will be queued anyway. This makes the batched update for isOffline and deletedAt simpler since there'll be no values that are unique to each asset.

Never thought of that, I've implemented your suggestion. I'm also considering changing the initial import code to ignore file mtime, this allows us to not do any file system calls except for the crawl. Metadata extraction will have to do the heavy lifting instead

Would that mean you queue them for metadata extraction even if they're unchanged? You can test it but I think it'd be more overhead than the stat calls.

Edit: also if you do this with the source set to upload, it would definitely be worse because it would queue a bunch of other things after metadata extraction.

@etnoy
Copy link
Contributor Author

etnoy commented Dec 12, 2024

The update to fileCreatedAt, fileModifiedAt and originalFileName is unnecessary and can be handled in metadata extraction since this will be queued anyway. This makes the batched update for isOffline and deletedAt simpler since there'll be no values that are unique to each asset.

Never thought of that, I've implemented your suggestion. I'm also considering changing the initial import code to ignore file mtime, this allows us to not do any file system calls except for the crawl. Metadata extraction will have to do the heavy lifting instead

Would that mean you queue them for metadata extraction even if they're unchanged? You can test it but I think it'd be more overhead than the stat calls.

Edit: also if you do this with the source set to upload, it would definitely be worse because it would queue a bunch of other things after metadata extraction.

I was referring to new imports, files that are new to immich. I hoped to improve the ingest performance by removing the stat call. After testing, there are two issues:

  • assetRepository.create requires mtime, which we can only get from stat. We could work around that by setting it to new Date(), but ideally it should be undefined
  • We still check for the existence of a sidecar, and this complicates things

If we can mitigate the two issues above, I can rewrite the library import feature and do that in batches as well!

@mertalev
Copy link
Contributor

mertalev commented Dec 12, 2024

I don't see why fileModifiedAt needs a non-null constraint in the DB. Might just be an oversight that didn't matter because it didn't affect our usage. I think you can change the asset entity and generate a migration to remove that constraint.

For sidecar files, maybe you could add.xmp to the glob filter and enable the option to make the files come in sorted order? That way you could make sure they're in the same batch.

@etnoy
Copy link
Contributor Author

etnoy commented Dec 12, 2024

I don't see why fileModifiedAt needs a non-null constraint in the DB. Might just be an oversight that didn't matter because it didn't affect our usage. I think you can change the asset entity and generate a migration to remove that constraint.

For sidecar files, maybe you could add.xmp to the glob filter and enable the option to make the files come in sorted order? That way you could make sure they're in the same batch.

I might just put new Date() in at the moment to keep the PR somewhat constrained.

Regarding sidecars, I have thought about that, problem right now is that we're batching the crawled files in batches of 10k. It might be hard to do get that working alright. Maybe I'll just queue a sidecar discovery for every imported asset for now

@mertalev
Copy link
Contributor

I think queueing sidecar discovery would introduce a race condition where it could run before, during or after metadata extraction. Since the refresh logic is already so much better, maybe leave the import for another PR so we can think about it more.

@etnoy etnoy force-pushed the feat/inline-offline-check branch from f980219 to 8944a32 Compare December 12, 2024 23:14
@etnoy etnoy force-pushed the feat/inline-offline-check branch from 8944a32 to 96f2f65 Compare December 12, 2024 23:16
@zackpollard
Copy link
Contributor

zackpollard commented Dec 17, 2024

I think queueing sidecar discovery would introduce a race condition where it could run before, during or after metadata extraction. Since the refresh logic is already so much better, maybe leave the import for another PR so we can think about it more.

Would this matter, because won't sidecar discovery re-queue metadata extraction afterwards if it does discover a sidecar file? I don't recall 100% if this is the behavior

@mertalev
Copy link
Contributor

I haven't looked closely at the sidecar discovery job either, but I think it's an issue either way since it's possible for jobs dependent on the metadata to behave differently. For example, a sidecar file that changes the orientation of the image won't be respected during thumbnail generation.

@etnoy etnoy force-pushed the feat/inline-offline-check branch 2 times, most recently from bbf20ab to d800c70 Compare December 18, 2024 00:15
@etnoy etnoy force-pushed the feat/inline-offline-check branch from d800c70 to 845b3f7 Compare December 18, 2024 00:24
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants