Bug: Slow compactor cleanups can cause consistency check errors on queriers

[jhalterman]

What is the bug?

Queriers currently provide a 45m grace period (3 * 15m bucket-store.sync-inverval) by default for store gateways to discover blocks before they can be queried. But the grace period is based on when the block was uploaded, not when it became discoverable via a bucket index. When bucket indexes don't get updated quickly, this can cause a block to be queryable before store gateways have had an opportunity to discover it. In a worst case scenario where a bucket index isn't updated for 45m, any new blocks become immediately queryable once the bucket index is loaded by a querier.

Proposed fix

Queriers should base their grace period, in whole or part, on when the bucket index was created, since that's when a block is discoverable by store gateways. For example, we could use 2 * the bucket-store.sync-inverval as the grace period relative to the bucket index creation time, but we could also (optionally) require that the existing 3 * the bucket-store.sync-inverval grace period relative to block upload time be satisfied, which would give a bit more time for store gateways to discover blocks in some cases.

How to reproduce it?

This came up when heavy OOO block creation caused a lot of compaction activity which slowed down bucket cleanup runs, leading to delayed bucket index creations.

What did you think would happen?

err-mimir-store-consistency-check-failed should not occur.

What was your environment?

K8s

Any additional context to share?

No response

[charleskorn]

A couple of thoughts on this:

• How would a querier know when a block first appeared in the bucket index? Would it need to examine all historical bucket index versions to discover this? Or would compactors set the time in the bucket index?
• Should we do the same thing for deletion times for blocks in the bucket index? Otherwise queriers could stop querying a compacted block (based on its deletion time, not when the deletion first appeared in the bucket index) well before the replacement block is required to be queried (based on when it first appeared in the bucket index, which may be well after the original block was deleted) - see Fix issue where both recently compacted blocks and their source blocks can be skipped during querying #9224 as an example of what I mean here

[jhalterman]

How would a querier know when a block first appeared in the bucket index? Would it need to examine all historical bucket index versions to discover this? Or would compactors set the time in the bucket index?

Having to read multiple versions of a bucket index doesn't sound good. Compactors including an approximate time that each block was indexed in the index itself would lighten the burden for queriers. This could probably even (eventually) replace the existing updated_at field that's in the index, which might not be needed anymore.

Another alternative is the block metas could be updated to include the time at which each block was first indexed, but that's potentially a lot of writes.

Should we do the same thing for deletion times for blocks in the bucket index?

Yes, possibly. Do you think that's worthy of a separate issue?