[
https://issues.apache.org/jira/browse/DRILL-5808?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16178361#comment-16178361
]
ASF GitHub Bot commented on DRILL-5808:
---------------------------------------
GitHub user paul-rogers opened a pull request:
https://github.com/apache/drill/pull/958
DRILL-5808: Reduce memory allocator strictness for "managed" operators
The "managed" external sort and the hash agg operators now actively attempt
to stay within a memory "budget."
Out goals are to:
1. Stay within the budget, and
2. Make full use of the available memory.
Unfortunately, at present, Drill has a number of limitations that work at
cross-purposes to the above goal.
* Upstream operators create record batches potentially larger than the
memory budget.
* Memory allocations are "lumpy" - power of two rounded.
* Vectors double in size automatically when needed.
The combination of the above means that memory planning must be aware of
the size of each and every vector to the byte level in order to predict size
doubling and power-of-two rounding.
But, of course, Drill is schema-on-read, meaning that Drill cannot know
ahead of time the "shape" of the data it will process. Without that
information, memory estimates are, at best, averages, but actual allocations
have a wide variance around those averages.
Add to this Drill's memory allocation scheme: each operator is given a
strict budget enforced by the memory allocator. Go above the budget by a single
byte and the query dies.
How do we resolve this conflict? On the one hand, Drill's internals are
rough-and-ready; it is impossible to predict actual memory usage. On the other
hand, the allocator requires perfect prediction else the user suffers with
failed queries.
Much work is needed in Drill internals to provide for better memory
management. (Relational databases have long ago solved the issues, so solutions
are available.) Until then, this commit introduces a work-around.
Memory-managed operators can ask for "leniency" from the allocator. In this
mode, the allocator:
* Allows actual memory use to spike up to 100% of the limit, or 100 MB,
whichever is less,
* Logs each such "excess allocation" as a warning, so we can identify and
fix issues, and
* Allows leniency only in production environments, but not during
development or test.
That is, we give users a margin for error so that their queries succeed
even if Drill's memory calculations don't come out exactly right.
This should be fine because, of course, Drill still has several operators
that observe no memory limits at all. Seems silly to have one operator using
GBs of memory, while enforcing a typical 30 MB limit on others.
Until all operators are memory managed, and Drill provides better memory
management tools, this PR allows queries to succeed even if we get things
slightly wrong internally.
You can merge this pull request into a Git repository by running:
$ git pull https://github.com/paul-rogers/drill DRILL-5808
Alternatively you can review and apply these changes as the patch at:
https://github.com/apache/drill/pull/958.patch
To close this pull request, make a commit to your master/trunk branch
with (at least) the following in the commit message:
This closes #958
----
commit a9c5083b8743efa2b5c74fee77e12d8f69258601
Author: Paul Rogers <prog...@maprtech.com>
Date: 2017-09-24T19:51:43Z
DRILL-5808: Reduce memory allocator strictness for "managed" operators
----
> Reduce memory allocator strictness for "managed" operators
> ----------------------------------------------------------
>
> Key: DRILL-5808
> URL: https://issues.apache.org/jira/browse/DRILL-5808
> Project: Apache Drill
> Issue Type: Improvement
> Affects Versions: 1.11.0
> Reporter: Paul Rogers
> Assignee: Paul Rogers
> Fix For: 1.12.0
>
> Original Estimate: 24h
> Remaining Estimate: 24h
>
> Drill 1.11 and 1.12 introduce new "managed" versions of the sort and hash agg
> that enforce memory limits, spilling to disk when necessary.
> Drill's internal memory system is very "lumpy" and unpredictable. The
> operators have no control over the incoming batch size; an overly large batch
> can cause the operator to exceed its memory limit before it has a chance to
> do any work.
> Vector allocations grow in power-of-two sizes. Adding a single record can
> double the memory allocated to a vector.
> Drill has no metadata, so operators cannot predict the size of VarChar
> columns nor the cardinality of arrays. The "Record Batch Sizer" tries to
> extract this information on each batch, but it works with averages, and
> specific column patterns can still throw off the memory calculations. (For
> example, having a series of very wide columns for A-M and very narrow columns
> for N-Z will cause a moderate average. But, once sorted, the A-M rows, and
> batches, will be much larger than expected, causing out-of-memory errors.)
> At present, if an operator is wrong in its memory usage by a single byte, the
> entire query is killed. That is, the user pays the death penalty (of queries)
> for poor design decisions within Drill. This leads to a less-than-optimal
> user experience.
> The proposal here is to make the memory allocator less strict for "managed"
> operators.
> First, we recognize that the managed operators do attempt to control memory
> and, if designed well, will, on average hit their targets.
> Second, we recognize that, due to the lumpiness issues above, any single
> operator may exceed, or be under, the configured maximum memory.
> Given this, the proposal here is:
> 1. An operator identifies itself as managed to the memory allocator.
> 2. In managed mode, the allocator has soft limits. It emits a warning to the
> log when the limit is exceeded.
> 3. For safety, in managed mode, the allocator enforces a hard limit larger
> than the configured limit.
> The enforcement limit might be:
> * For memory sizes < 100MB, up to 2x the configured limit.
> * For larger memory sizes, no more than 100MB over the configured limit.
> The exact numbers can be made configurable.
> Now, during testing, scripts should look for over-memory warnings. Each
> should be fixed as we fix OOM issues today. But, during production, user
> queries are far less likely to fail due to any remaining corner cases that
> throw off the memory calculations.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
