Project

General

Profile

Actions

Idea #8886

closed

[API] Do not block individual API queries on rebuilding the permissions graph

Added by Brett Smith about 8 years ago. Updated almost 8 years ago.

Status:
Resolved
Priority:
Normal
Assigned To:
Category:
API
Target version:
Start date:
04/13/2016
Due date:
Story points:
2.0

Description

Functional requirements:

  • Do not change the database structure.
  • Do not write new Go code. As much as possible, we want to use the existing code, just adapt it to run out of band from the request cycle.
  • Updating the permissions cache happens outside the normal request handler cycle. No individual request waits for the permissions cache to be updated before it is handled. Each request just uses the "latest" permissions cache available, at least from the first time that it needs the cache.
    Unless it is unreasonably hard, each request should use the same permissions cache throughout its lifecycle. (In general we want this, and if it's low-hanging fruit, this seems like a good opportunity to make that happen. However, we believe that we do not currently have this behavior today, and fixing this is not so pressing that we're willing to hold up the performance improvement on it.)
  • This new permissions cache handling is behind a feature flag in application.yml called async_permissions_update. When the flag is falsy, the API server continues to use its current behavior of rebuilding the permissions cache when needed. The default in application.default.yml is false.

Subtasks 3 (0 open3 closed)

Task #9090: Write permission rebuild serviceResolvedPeter Amstutz04/13/2016Actions
Task #8978: ReviewResolvedPeter Amstutz04/13/2016Actions
Task #9091: TestingClosedPeter Amstutz04/13/2016Actions

Related issues

Related to Arvados - Bug #8787: [API] repositories/get_all_permissions method is too slowResolved03/24/2016Actions
Precedes (1 day) Arvados - Idea #9186: [API] Test client impact of async_permissions_update=TrueClosed04/15/201604/15/2016Actions
Precedes (1 day) Arvados - Idea #9502: [API] Update permissions cache as needed after select writesResolved06/28/2016Actions
Actions #1

Updated by Brett Smith about 8 years ago

  • Description updated (diff)
  • Category set to API
Actions #2

Updated by Brett Smith about 8 years ago

  • Subject changed from [API] Do not block the handling of individual API queries on rebuilding the permissions graph to [API] Do not block individual API queries on rebuilding the permissions graph
Actions #3

Updated by Brett Smith about 8 years ago

  • Target version set to Arvados Future Sprints
Actions #4

Updated by Brett Smith about 8 years ago

Brett Smith wrote:

  • Does Rails provide an easy way to do this through its API? A sort of atomic "replace cached result" method?

That seems unlikely to meet the requirements. Instead the basic idea is we want to find the cheapest possible way to do something "in the background" (i.e., outside the request cycle) for a Rails server.

  • Should we move to eventual consistency?

Yes. At the very least, this seems like a good opportunity to experiment with it and see if clients break to think about whether or not we want to have more eventual consistency throughout the API.

  • Should we have a separate background process that listens to the relevant tables and builds a computed table with simple (subject, has_permission_to, object) tuples?

Not if we can help it, at this time. We want the smallest possible implementation that can work for now.

  • Should we require a newer version of PostgreSQL that can traverse the permissions graph for us (in C)?
    • This may require adjustments to the way we store permissions in the database generally.

This seems like an implementation detail. A worthy one, but still. First we should get the process out of the request cycle. Once we do that, this seems like a potential optimization.

  • It would be nice, but not critical, if this was behind a feature flag. That way we could test it on clusters that we know have large permissions graphs before deploying it more broadly.

We do want it behind a feature flag, because we want to enable this on a single "real" cluster (e.g., not 4xphq or c97qk) without enabling it on other production clusters. We also want to be able to roll it back quickly if it turns out not to work well.

Actions #5

Updated by Brett Smith about 8 years ago

  • Description updated (diff)
Actions #6

Updated by Tom Clegg almost 8 years ago

Possible approach: Run a separate service that updates the permission graph as needed.

Service pseudocode:
  • LISTEN for relevant changes on groups table
  • while true:
    • Let new_graph_timestamp = db_current_time
    • Compute new_graph
    • Write new_graph to Rails cache
    • Write new_graph_timestamp to Rails cache
    • Wait for the database to notify about a change to the groups table
Presumably, many test cases expect permissions to be applied right away. We will need to fix them by adding a "wait for all permission updates to finish" mechanism. This seems like a useful API to have in the real server, too. The relevant condition is
  • The permission graph has been updated since time T (e.g., the modification time found in an API response), and
  • The permission service is in the "wait for NOTIFY" state, and
  • There are no Postgres notifications pending

For purposes of test cases, we could have a synchronous "recompute permission graph" API, with guards similar to the "database reset" API.

Actions #7

Updated by Tom Clegg almost 8 years ago

  • Target version changed from Arvados Future Sprints to 2016-05-11 sprint
Actions #8

Updated by Peter Amstutz almost 8 years ago

  • Assigned To set to Peter Amstutz
Actions #9

Updated by Peter Amstutz almost 8 years ago

  • Story points set to 2.0
Actions #10

Updated by Tom Clegg almost 8 years ago

Should remove the unused copy-and-paste stuff like dispatch_argv and "...the logs table, then push the notification onto the EventMachine channel".

If we get 10x "notify" in a row, do we recalculate the permissions 10x? Postgres says "notifications from different transactions will never get folded into one notification" which seems to mean we need to dedup them ourselves, otherwise it seems like we can easily end up with a backlog of useless recalculations based on notifications that are already older than our cache. The new_graph_timestamp idea from note-6 might help with that.

In permission-updater, the LISTEN statement should be issued before the initial calculation so we don't miss notifications.

If the rails cache expires (or just hasn't been populated yet) is there some reasonably easy way for group_permissions to notify the updater (in case it's "cache expired") and then wait for the updater to finish? This should avoid stuff like "server just came up, and now has lots of threads all trying to recompute the permission graph for themselves, which is even slower because there are so many threads doing it"...

Actions #11

Updated by Peter Amstutz almost 8 years ago

Tom Clegg wrote:

Should remove the unused copy-and-paste stuff like dispatch_argv and "...the logs table, then push the notification onto the EventMachine channel".

Yep, fixed.

If we get 10x "notify" in a row, do we recalculate the permissions 10x? Postgres says "notifications from different transactions will never get folded into one notification" which seems to mean we need to dedup them ourselves, otherwise it seems like we can easily end up with a backlog of useless recalculations based on notifications that are already older than our cache. The new_graph_timestamp idea from note-6 might help with that.

Added last_updated_permissions which would help with that.

In permission-updater, the LISTEN statement should be issued before the initial calculation so we don't miss notifications.

Done.

If the rails cache expires (or just hasn't been populated yet) is there some reasonably easy way for group_permissions to notify the updater (in case it's "cache expired") and then wait for the updater to finish? This should avoid stuff like "server just came up, and now has lots of threads all trying to recompute the permission graph for themselves, which is even slower because there are so many threads doing it"...

Now it polls the cache until it gets a non-nil result. It's a bit of a hack but it seems to work.

Actions #12

Updated by Tom Clegg almost 8 years ago

at e711e48...

The comments in user.rb should be updated: the explanation of how the graph is built should be on calculate_group_permissions, not the method that just returns a cached result.

All test suites are failing when test_helper calls invalidate_permissions_cache.

Presumably the group_permissions method still needs to call calculate_group_permissions if async_permissions_update is false, or something... otherwise the default configuration will just hang?

Actions #13

Updated by Tom Clegg almost 8 years ago

9e17d8d... LGTM thanks

Actions #14

Updated by Peter Amstutz almost 8 years ago

  • Status changed from New to In Progress
Actions #15

Updated by Brett Smith almost 8 years ago

  • Status changed from In Progress to Resolved

Now that the code specified by the story is done and merged, #9186 is the follow-up story to test it and see how viable it is in a production-like deployment.

Actions

Also available in: Atom PDF