TransactionClient gc function often enters error loop during CleanupLocks

[jmhrpr]

I have a program that runs periodically which makes a single call to the TransactionClient gc function. Around half of the time it enters some loop which prints thousands of these logs and eventually OOMs. It is checking many ranges/regions for the the same key. It is a different key each time it enters the loop. They are INFO logs originating from client-rust-da362376b56921db/1fa846b/src/request/plan.rs:686.

CleanupLocks::execute, inner region error:key <REDACTED KEY> is not in region key range [<REDACTED RANGE A START>, <REDACTED RANGE A END>) for region 116477
CleanupLocks::execute, inner region error:key <REDACTED KEY> is not in region key range [<REDACTED RANGE B START>, <REDACTED RANGE B END>) for region 116288
CleanupLocks::execute, inner region error:key <REDACTED KEY> is not in region key range [<REDACTED RANGE C START>, <REDACTED RANGE C END>) for region 970120
CleanupLocks::execute, inner region error:key <REDACTED KEY> is not in region key range [<REDACTED RANGE D START>, <REDACTED RANGE D END>) for region 969086
...

[pingyu]

Could you show your program or some snippet can reproduce this issue ?

As well as the setup of TiKV cluster.

[ekexium]

I'm afraid GC is not stable yet(#180).
After a quick look at the code (haven't reviewed this part for years), one possible problem to investigate: the scan_lock request might be directly passed to cleanup_locks without proper region setup (e.g. via something like retry_multi_region)

client-rust/src/transaction/client.rs

Lines 264 to 271 in 59f13b5

	let req = new_scan_lock_request(range, safepoint, options.batch_size);
	let plan = crate::request::PlanBuilder::new(self.pd.clone(), self.keyspace, req)
	.cleanup_locks(ctx.clone(), options, backoff, self.keyspace)
	.retry_multi_region(DEFAULT_REGION_BACKOFF)
	.extract_error()
	.merge(crate::request::Collect)
	.plan();
	plan.execute().await

This change was introduced in PR #378, which added async commit lock resolution support. As a workaround, you might want to temporarily disable async commit and use the pre-PR GC implementation until this issue is resolved.