CI Failure Failed to allocated 4870880 bytes in ManyPartitionsTest

[ballard26]

Child issue of #7405

After replicating the issue a few times on x86 and ARM with --dump-memory-diagnostics-on-alloc-failure-kind=all the issue isn't Redpanda running out of memory. Rather that a ~4Mb contiguous chunk of memory couldn't be allocated due to fragmentation. See;

Used memory:  1267M
Free memory:  3733M
Total memory: 5G
...
ERROR 2023-01-18 21:11:50,914 [shard  1] seastar - Failed to allocate 4870880 bytes

Links to various failures:

https://buildkite.com/redpanda/vtools/builds/4913#0185500b-bdaa-4cb0-b640-31259ecac3be
https://buildkite.com/redpanda/vtools/builds/4918#0185552f-c829-43a8-9d3b-b74db0421f55
https://buildkite.com/redpanda/vtools/builds/4898#018545bd-095d-4b37-b549-0173b1d272c1

Decoded backtrace for the failure:

[brandonallard@fedora temp]$ llvm-addr2line --obj b80b71a6b9978d8a9767049adf4e973b443f35.debug -f  --demangle  --pretty-print 
0x5656c75
seastar::memory::on_allocation_failure(unsigned long) at /v/build/v_deps_build/seastar-prefix/src/seastar/src/core/memory.cc:1806
0x5665351
operator new(unsigned long) at /v/build/v_deps_build/seastar-prefix/src/seastar/src/core/memory.cc:2064
0x2b25d22
void* std::__1::__libcpp_operator_new<unsigned long>(unsigned long) at /vectorized/llvm/bin/../include/c++/v1/new:245
0x2b3200d
kafka::handler_template<kafka::metadata_api, (short)0, (short)7, seastar::future<seastar::foreign_ptr<std::__1::unique_ptr<kafka::response, std::__1::default_delete<kafka::response>>>>, &kaf
ka::metadata_memory_estimator(unsigned long, kafka::connection_context&)>::handle(kafka::request_context, seastar::smp_service_group) (.resume) at /var/lib/buildkite-agent/builds/buildkite-a
md64-builders-i-02c556a1c5e374299-1/redpanda/redpanda/src/v/kafka/server/handlers/metadata.cc:420

Analysis of the failure

get_topic_metadata in src/v/kafka/server/handlers/metadata.cc will allocate a std::vector<metadata_response_partition> for every topic and sizeof(metadata_response_partition) == 112 . In the ManyPartitionsTest test we create a topic with 43490 partitions. Hence we are trying to allocate exactly (112*43490=)4870880 contiguous bytes to store metadata for every partition in the test topic.

Potential fix

The best bet is to switch these response types to use a fragmented vector if possible. That way we are avoiding large contiguous allocations. Another potential issue to look out for is when we encode the response. If the encoder tries to dynamically allocate a contiguous chunk of memory to encode the vector we could run into the same issue.

[dotnwat]

thanks @ballard26 we'll take a looksie

[piyushredpanda]

@ballard26 and @travisdowns chatted about this one and Brandon will help move to fragmented vector here. Thanks for taking it up, @ballard26!