update abseil

[kyessenov]

Signed-off-by: Kuat Yessenov kuat@google.com

Description: Update abseil library to 5/17
Risk Level: Low
Testing: No extra
Docs Changes: None
Release Notes:

[kyessenov]

cc @g-easy : trying to divide and conquer your PR.

[g-easy]

Thanks! :)

[kyessenov]

This comes with a fix to use node_hash_map instead of flat_hash_map. The difference is that the former does not move the values on the table re-organization. This prevents the double destructor call which appears to cause the ASAN violation. I think the upgrade of absl changed the heuristics in the hash map, so it re-orged in the unit test. That doesn't mean the error is absent on master, it just revealed itself after the upgrade.

/cc @lizan

[kyessenov]

I'm not sure how to deal with ipv6 test. I'm getting it to pass locally:

bazel test --test_env=ENVOY_IP_TEST_VERSIONS=v6only //test/integration:stats_integration_test

[lizan]

@kyessenov what's the version of libstdc++ in your local? The m_per_cluster is different across environments, so just set it what CI says, not your local one.

[lizan]

This comes with a fix to use node_hash_map instead of flat_hash_map. The difference is that the former does not move the values on the table re-organization. This prevents the double destructor call which appears to cause the ASAN violation. I think the upgrade of absl changed the heuristics in the hash map, so it re-orged in the unit test. That doesn't mean the error is absent on master, it just revealed itself after the upgrade.

Nice catch, it is worth to submit that part as separate PR, with a regression test.

[kyessenov]

@lizan

How do I satisfy both versions of libstdc++? I see that the regular test run matches my local environment and expects 49193 in the test. But then IPv6 run must be using a different environment and expects a different number 49137. Do you want me to put both numbers in the test?

[lizan]

@kyessenov I think you should only put 49137, which should pass all CI. That test won't pass in local environments.

[kyessenov]

Is there anything I can do to pass macos build? I don't have a mac so not sure how to debug the CI failure.

[lizan]

It seems failing with thread annotation, @jmarantz do you have any idea from abseil changes?

[jmarantz]

I do not know why this would start failing now. The thread annotation is a function of clang more than absl.

[jmarantz]

nice reduction of almost 1k per cluster. I'm surprised absl made that much of a difference. I assume this is not because o of the service node_hash_map, which I don't think would be replicated per cluster.

[kyessenov]

The reduction was before I made node_hash_map change I think.

[kyessenov]

Is there anything I can do to debug macos build? I don't have a mac laptop. Should I borrow it from someone to reproduce locally?

[lizan]

@kyessenov I'm working on a workaround, do you mind me take over this PR? Upgrading absl to more recent version and use absl thread annotation would solve it.

[kyessenov]

Sure, go ahead.

[kyessenov]

Superseded by #7199