Tablet unload impacted by long-running compaction cancellation

[dtspence]

Describe the bug
A tablet unload (i.e. due to migration request) may be delayed while a tablet is attempting to unload, but cannot due to pending compaction cancellations. We have observed that the tablet will wait for as long as 50+ minutes while compactions cancel.

Versions (OS, Maven, Java, and others, as appropriate):

• Affected version(s) of this project: 2.1.2

To Reproduce
We are attempting to gather additional information to reproduce. Some preliminary information:

• Tablets are being compacted by t-servers (i.e. not using external compactions).
• Data being compacted is not expected to be filtered, however we are unsure if somehow the iterator may not be returning for an extended time.

Expected behavior
Migration request should complete within some shorter time.

Screenshots
N/A

Additional context
The manager logs:

2024-04-22T16:23:58,434 [balancer.HostRegexTableLoadBalancer] WARN: Not balancing tables due to 1 outstanding migrations

[dlmarion]

@dtspence - is it the case that you manually cancelled the compactions? If so, did your command complete, or did it hang too?

[dlmarion]

The FileCompactor checks if the compaction is still enabled for every key that it writes. I'm curious if the compaction was making progress (you said filtering was not expected which could also be a cause). Is this happening often? If not, is bouncing the tserver an option?

[dtspence]

@dlmarion

is it the case that you manually cancelled the compactions? If so, did your command complete, or did it hang too?

No, the compaction eventually logs that it has canceled. We have not been taking manual intervention.

The FileCompactor checks if the compaction is still enabled for every key that it writes. I'm curious if the compaction was making progress (you said filtering was not expected which could also be a cause). Is this happening often? If not, is bouncing the tserver an option?

Yes, the issue is re-appearing. It does not appear to be localized to a single t-server. We have been wondering if something tablet related that correlates to the issue. At least one tablet we were looking at was a hot-spot and contained a lot of i-files from imports.

[dlmarion]

@dtspence - are you seeing a lot of these messages?

[dtspence]

@dlmarion

are you seeing a lot of these messages?

No we are not seeing the message above. Just for reference, it may have already been known - but we see:

2024-04-23T17:58:17,298 [tserver.UnloadTabletHandler] DEBUG: Failed to unload tablet <tablet-name> ... it was already closing or closed

[dlmarion]

I think that log message might be from the Manager continuing to tell the TabletServer to unload the tablet.

[dlmarion]

I'm still thinking that maybe the compaction is not making progress. I don't think there is good logging for this with compactions that run in the Tablet Server. IIRC, the way to tell if it's making progress is to check out the output file for the compaction in HDFS and see if its size is increasing. If nothing is getting written to the file for a long time, then either it's filtering out a lot of data, or it's waiting on input from HDFS.

[ivakegg]

I would love to have something in place to avoid compaction from holding up unloading of tablets. Is this something that is relative easy to do? This would save us from long shutdowns as well.

[dlmarion]

Is this something that is relative easy to do?

It's not a switch that exists today. We would need to develop and test a solution. If you can identify which compactions are causing a tablet not to close, then you could run them as External Compactions. The existing codepath does not wait for the External Compactions to complete. It only waits for them if they are in the process of committing their changes to the tablet metadata.

[dtspence]

We believe the compaction is not emitting a key/value to respond to the cancellation. We also observed that the issue appears to occur when a tablet splits (due to growth).

In all (or most) of the observed instances the following occurs:

• T-server performs a sized-based split on a tablet into A and B.
• T-server starts compaction for A.
• Manager begins to migrate A and asks the t-server to close A.
• T-server attempts to unload tablet and cancel the running compaction(s).
• T-server eventually cancels compactions and tablet is unloaded.

[dlmarion]

If we introduced a delay in starting compactions for newly split tablets, then that might help mitigate this issue. We would have to make the delay time configurable most likely as timing the balancer can't really be done.

[dlmarion]

If we introduced a delay in starting compactions for newly split tablets, then that might help mitigate this issue. We would have to make the delay time configurable most likely as timing the balancer can't really be done.

In 2.1, Tablet.split() is the only caller of the TabletData constructor that sets Tablet.splitCreationTime. In a normal Tablet load scenario, splitCreationTime remains zero. We could use this variable to introduce a delay in starting compactions for newly split tablets considering there is a good chance that at least one of them will be migrated.