On June 17, I updated the crate version from 5.10.7 (Stable) to 5.10.9 (testing).
Regarding the Java heap space issue, I observed a single behavioral difference.
Using several group by clauses in my query used to throw circuit-breaker exception in the logs. It now takes a crate node down and displays a Java heap space issue.
Does version 5.10.9 include any changes related to this behavior?
There was a change related to CircuitBreakerException in 5.10.8, which introduced a regression fixed in 5.10.9, but nothing that would explain what you are seeing, at least at a quick glance.
Could you please share a query so I can try to reproduce?
I tried with several datasets and I always got CircuitBreakingException as expected, there must be something about your data/setup what circumvents ram accounting.
CrateDB’s ram accounting is best-effort, since accurately predicting how much memory a query will take is extremely hard. We use thresholds.
Does this only happen in a table with 100s of millions? or does it also happens in other ‘big’ tables, If so, could you share a table definition that triggers your issue? SHOW CREATE TABLE <table_name>
If column names are sensitive please anonymize them.