redpanda
redpanda copied to clipboard
redpanda-data
tx/compaction: reduce memory usage with lots of aborted transactions
Who is this for and what problem do they have today?
With the current compaction implementation for transactions, we cache aborted transactions for a segment in memory. In a pathological case where the # of aborted tx >> # non aborted ranges, we can reduce memory usage by only caching the working set of aborted tx ranges.
What are the success criteria?
Reduced memory usage with existing correctness guarantees.
Why is solving this problem impactful?
Reduce memory usage.
