Failure to compact due to maximum index size 64 GiB

[thejosephstevens]

Describe the bug
Over the weekend we significantly expanded one of our clusters, pushing ~153M timeseries to our Cortex 1.11.1 cluster in a day.

To Reproduce
Steps to reproduce the behavior:

1. Start Cortex 1.11.1 in single-tenant mode as microservices
2. Push ~153M timeseries to it
3. See the Compactor get to Stage 3 compaction and then fail with an error like this

level=error ts=2022-04-04T16:11:41.167458322Z caller=compactor.go:537 component=compactor msg="failed to compact user blocks" user=fake err="compaction: group 0@5679675083797525161: compact blocks [data/compact/0@5679675083797525161/01FZNK9ZNWHJWRXB2YNHZ9W8H9 data/compact/0@5679675083797525161/01FZPV9DZ1PGV0HHV9007P8ZWM]: \"data/compact/0@5679675083797525161/01FZTCHWRJHFN419FJPDM0MQNW.tmp-for-creation/index\" exceeding max size of 64GiB"

The two blocks referenced by this error are 12-hour blocks at level-3 compaction, each of which has an index of ~38 GB (summing together to ~76 GB > 64 GB).

Expected behavior
There should be a way of skipping this, forcing compaction, sharding, or something.

There's an upstream Thanos patch here which allows skipping compaction but it appears to not be used by Cortex today (found by @alvinlin123 in Cortex Slack). There's also a Thanos change here which would automatically skip past compaction if the block is too large.

Mimir appears to have the ability to get past this by sharding during compaction so multiple blocks are produced per day, each of which can have a smaller index.

Environment:

• Infrastructure: Kubernetes (GKE)
• Deployment tool: Jinja (derived from the cortex-jsonnet manifest)

Storage Engine

• Blocks
• Chunks

Additional Context

[alvinlin123]

There is a related Prometheus issue to support index file size bigger than 64Gi

However, I think we shouldn't wait for that, and should have Cortex skip compaction for blocks with large index.

[alvinlin123]

After PR #4707 I still need to implement the part that auto skip compaction for blocks with humongous index.

[alvinlin123]

Hmm looks like I can't just use Thanos largeTotalIndexSizeFilter for Cortex's ShuffleShardingPlanner because ShuffleShardingPlanner is coupled with non-exported tsdbBasedPlanner struct.

[alvinlin123]

Another relevant Thanos issue thanos-io/thanos#3068 to track the sharding work

[stale]

This issue has been automatically marked as stale because it has not had any activity in the past 60 days. It will be closed in 15 days if no further activity occurs. Thank you for your contributions.

[thejosephstevens]

For what it's worth, we migrated to Mimir and no longer have issues with this. I think this still makes sense to implement in Cortex since it's a pretty brutal issue to hit (and effectively puts a hard cap on max tenant size), but it doesn't need to be kept open for us. If you'd like to close it for clean-up, feel free.

[jeromeinsf]

@thejosephstevens we are looking on how to address this in Cortex.

[yeya24]

@alvinlin123 I don't think we can close this issue since only the proposal gets merged. Please reopen it.