Boaz Ben-Zvi created DRILL-6543:
-----------------------------------
Summary: Options for memory mgmt: Reserve allowance for
non-buffered, and Hash-Join default to not fallback
Key: DRILL-6543
URL: https://issues.apache.org/jira/browse/DRILL-6543
Project: Apache Drill
Issue Type: Improvement
Components: Execution - Relational Operators
Affects Versions: 1.13.0
Reporter: Boaz Ben-Zvi
Assignee: Boaz Ben-Zvi
Fix For: 1.14.0
Changes to options related to memory budgeting:
(1) Change the default for "drill.exec.hashjoin.fallback.enabled" to *false*
(same as for the similar Hash-Agg option). This would force users to calculate
and assign sufficient memory for the query, or explicitly choose to fallback.
(2) When the "planner.memory.max_query_memory_per_node" (MQMPN) option is set
equal (or "nearly equal") to the allocated *Direct Memory*, an OOM is still
possible. The reason is that the memory used by the "non-buffered" operators is
not taken into account.
For example, MQMPN == Direct-Memory == 100 MB. Run a query with 5 buffered
operators (e.g., 5 instances of a Hash-Join), so each gets "promised" 20 MB.
When other non-buffered operators (e.g., a Scanner, or a Sender) also grab some
of the Direct Memory, then less than 100 MB is left available. And if all those
5 Hash-Joins are pushing their limits, then one HJ may have only allocated 12MB
so far, but on the next 1MB allocation it will hit an OOM (from the JVM, as all
the 100MB Direct memory is already used).
A solution -- a new option to _*reserve*_ some of the Direct Memory for those
non-buffered operators (e.g., default %25). This *allowance* may prevent many
of the cases like the example above. The new option would return an error (when
a query initiates) if the MQMPN is set too high. Note that this option +can
not+ address concurrent queries.
This should also apply to the alternative for the MQMPN - the
{{"planner.memory.percent_per_query"}} option (PPQ). The PPQ does not
_*reserve*_ such memory (e.g., can set it to %100); only its documentation
clearly explains this issue (that doc suggests reserving %50 allowance, as it
was written when the Hash-Join was non-buffered; i.e., before spill was
implemented).
The memory given to the buffered operators is the highest calculated between
the MQMPN and the PPQ. The new reserve option would verify that this figure
allows the allowance.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
