Csaba Ringhofer created IMPALA-10430:
----------------------------------------
Summary: Allow limiting duplicate rows from joins
Key: IMPALA-10430
URL: https://issues.apache.org/jira/browse/IMPALA-10430
Project: IMPALA
Issue Type: Improvement
Components: Backend
Reporter: Csaba Ringhofer
Small mistakes in join conditions can lead to huge number of duplicates and
turn a simple query into a massive resource hog until it is cancelled (or runs
out of memory). An even worse case is if this happens in an INSERT and data is
being dumped to HDFS until some quote is hit.
An idea for a safety net is to track the number of "original" rows (that come
from scanning a table) and "joined" rows (that come from joins), and abort the
query if joined/original is above a certain ratio.
E.g. the following query options could control this:
max_joined_original_ratio
min_rows_to_enforce_joined_original_ratio (to allow intentional cross joins for
smaller datesets)
So the query could be cancelled if
max(max_joined_original_ratio * original,
min_rows_to_enforce_joined_original_ratio ) < joined
We could find default values that would rarely limit sane queries, e.g.
max_joined_original_ratio = 100,
min_rows_to_enforce_joined_original_ratio=10,000,000,000
If someone knows that the query should have no duplicates at all then this
could be enforced with
max_joined_original_ratio = 1,
min_rows_to_enforce_joined_original_ratio=0
Tracking 'original' and 'joined' could be done several ways, but it is
important that it should be done "globally", not per join node , because
consecutive joins' "duplicate factors" are multiplied, e.g. five joins that
all have a ratio of 10x would result in 10000x global ratio. A possibility is
to use profile counters in scan and join nodes and sum them in the coordinator.
Apart from safety from runaway queries this could also help the planner by
having enforced limits, especially in the max_joined_original_ratio = 1 case.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
