select * from mutation_fragments on tablets may miss data during tablet migration about scylladb

[tgrabiec]

c7908c3 contains a workaround for this bug by executing a read barrier after add_replica API so that query on the new replica is guaranteed to see the non-transitional erm. We should remove the workaround after this is fixed.

[denesb]

This was seen cause a test failure in CI: https://jenkins.scylladb.com/job/scylla-master/job/scylla-ci/8025/testReport/junit/(root)/non-boost%20tests/Tests___Unit_Tests___cql_pytest_test_select_from_mutation_fragments_release_2/
Adding ci-stability label.

[avikivity]

Although it causes a test failure, I consider it a minor bug. SELECT mutation_fragments() is weird in that it is node-local. If you do node-local things on things that can migrate, what do you expect?

[nyh]

Although it causes a test failure, I consider it a minor bug. SELECT mutation_fragments() is weird in that it is node-local. If you do node-local things on things that can migrate, what do you expect?

Exactly. We have an API to temporarily disable migration, so why not use it while trying to do node-local reads? In Python we can even have a context manager to make it easy to temporarily disable migration. It reminds me of code we had in OSv to run CPU-local things while disabling thread migration between CPUs :-)

[denesb]

Ok, I will send a PR tomorrow, to disable migrations for this entire test file.

Although... now that I think of it, the test is ran against a single node cluster and we don't have intra-node migrations yet. So what migrations could we have here? Tablet split?

[raphaelsc]

Ok, I will send a PR tomorrow, to disable migrations for this entire test file.

Although... now that I think of it, the test is ran against a single node cluster and we don't have intra-node migrations yet. So what migrations could we have here? Tablet split?

bumped into this locally

[denesb]

Ok, I will send a PR tomorrow, to disable migrations for this entire test file.
Although... now that I think of it, the test is ran against a single node cluster and we don't have intra-node migrations yet. So what migrations could we have here? Tablet split?

bumped into this locally

You mean the test failed?

[bhalevy]

This is a minor issue since SELECT * FROM MUTATION_FRAGMENTS is used for debugging, so can be postponed to 6.1.

@mykaul / @dani-tweig WDYT?

[dani-tweig]

This is a minor issue since SELECT * FROM MUTATION_FRAGMENTS is used for debugging, so can be postponed to 6.1.

@mykaul / @dani-tweig WDYT?

@bhalevy
I vote for 6.1 + maybe add a note in the release notes if you think people will encounter this symptom.

[mykaul]

This is a minor issue since SELECT * FROM MUTATION_FRAGMENTS is used for debugging, so can be postponed to 6.1.
@mykaul / @dani-tweig WDYT?

@bhalevy I vote for 6.1 + maybe add a note in the release notes if you think people will encounter this symptom.

No need for the release notes - it's a debugging aid.