tiered compaction: duplicated L1 layer error in test_deletion_queue_recovery

[arpad-m]

Running the test_deletion_queue_recovery or test_uploads_and_deletions tests with tiered compaction enabled gives "duplicated L1 layer" errors:

2024-05-10T22:25:57.275644Z ERROR request{method=PUT path=/v1/tenant/b8c7c9b6739fed9060bfdf938ec9e9dc/timeline/5ffef4897699e4ff0fd68add218821ea/checkpoint request_id=227fdbbe-247c-4aee-8545-524487816dc4}:manual_checkpoint{tenant_id=b8c7c9b6739fed9060bfdf938ec9e9dc shard_id=0000 timeline_id=5ffef4897699e4ff0fd68add218821ea}: duplicated L1 layer layer=000000067F00000005000000000000000001-030000000000000000000000000000000002__0000000001535489-000000000154E229-00000001
2024-05-10T22:25:57.275660Z ERROR request{method=PUT path=/v1/tenant/b8c7c9b6739fed9060bfdf938ec9e9dc/timeline/5ffef4897699e4ff0fd68add218821ea/checkpoint request_id=227fdbbe-247c-4aee-8545-524487816dc4}:manual_checkpoint{tenant_id=b8c7c9b6739fed9060bfdf938ec9e9dc shard_id=0000 timeline_id=5ffef4897699e4ff0fd68add218821ea}: duplicated L1 layer layer=000000067F00000005000040000000000001-030000000000000000000000000000000002__0000000001535489-000000000154E229-00000001

visible with the following diff of test_deletion_queue_recovery:

-    env = neon_env_builder.init_start(initial_tenant_conf=TENANT_CONF)
+    tenant_conf = TENANT_CONF
+    tenant_conf["compaction_algorithm"] = '{{"kind": "Tiered"}}'
+    env = neon_env_builder.init_start(initial_tenant_conf=tenant_conf)

The test_deletion_queue_recovery test ran into all the important issues: previously, it ran into #7244 and #7296.

part of #7554

Tasks

Beta Give feedback

1. Add test_uploads_and_deletions test #7758
   +0
2. fix the issue
3. remove the allowed_errors added in Add test_uploads_and_deletions test #7758

[jcsp]

The test_deletion_queue_recovery test ran into all the important issues: previously, it ran into #7244 and #7296.

Can we lift the subset of this test that reproduces these issues into a dedicated compaction test, perhaps as part of the PR fixing this issue?

[arpad-m]

Can we lift the subset of this test that reproduces these issues into a dedicated compaction test

I could file a PR and then just allow the duplicated L1 layer errors.

[arpad-m]

I could file a PR and then just allow the duplicated L1 layer errors.

Done: #7758

[problame]

Meeting notes:

• quite a serious condition: we throw away the second struct Layer but we overwrote the on-disk file
• => PS PageCache incoherency if the bit pattern of the new file is not identical
• need to investigate & fix this, it's a potential problem right now (if compact legacy is not bitpattern-deteterministic).
  • issue: re-assess severity of duplicate layers: nowadays it cannot happen & we should panic/abort() early if they do #7790
• But let's not get distracted => Goal for tiered compaction: never get into this situation in the first place.
• Arpad: assumes root cause is too many loop iterations (just speculation though)

Action item: arpad & heikki to understand why it happens.