Hey Udi,
Thanks for the commit comment
<https://github.com/apache/beam/commit/f08f21cdf4c067745a10b31a6481ed470f97dadc#r31216133>.
I'll try to dump any (old) mental context I have left..
We were trying to find the right point in a space of:
* enough parallelism to speed things up
- more than 256 didn't seem to help with perf, and 256 was not close to
looking like a DDOS to GCS, so we were not in danger of quota limits being
imposed.
* enough batches to speed things up
- batches -> fewer RPCs from the job to GCS, more RPCs inside GCS. 100
was again good for perf and good for quotas.
* small enough batches to spread the load
- if you think about the multiple layers of fanout in RPC handling,
sending a batch RPC cuts out the first layer of load-balancing. The
endpoint that receives the batch itself handles all the individual
requests, and if that endpoint is slow or any individual request in the
batch is slow, the entire batch is slow. Prefer many batches in flight, so
as to not be limited by the performance of that single endpoint.
Coming to the question in the PR comment:
> Any reason not to use this.executorService?
I think the main reason to not use `this.executor` is that we didn't have
constraints on that executor in terms of either upper or lower bounds on
parallelism, so it seemed safer to use our own with known limits.
Thanks for catching the memory leak though - we didn't at the time! I'll
defer to you (and especially Luke ;) on a good solution to fix.
Dan
On Thu, Nov 8, 2018 at 2:12 PM Udi Meiri <[email protected]> wrote:
> My thought was to use 1 executor per GcsUtil instance (or 1 per process as
> you suggest), with a possible performance hit since I don't know how often
> these batch copy and remove operations are called.
> The other option is to leave things as they mostly are, and only remove
> the call to getExitingExecutorService.
>
> Both options risk delaying worker shutdown if the executor's shutdown()
> hasn't been called, which is I guess why the executor in GcsOptions.java
> creates daemon threads.
>
> On Thu, Nov 8, 2018 at 1:02 PM Lukasz Cwik <[email protected]> wrote:
>
>> Not certain, it looks like we should have been caching the executor
>> within the GcsUtil as a static instance instead of creating one each time.
>> Could have been missed during code review / slow code changes over time.
>> GcsUtil is not well "loved".
>>
>> On Thu, Nov 8, 2018 at 11:00 AM Udi Meiri <[email protected]> wrote:
>>
>>> HI,
>>> I've identified a memory leak when GcsUtil.java instantiates a
>>> ThreadPoolExecutor (https://issues.apache.org/jira/browse/BEAM-6018).
>>> The code uses the getExitingExecutorService
>>> <https://github.com/apache/beam/blob/279a05604b83a54e8e5a79e13d8761f94841f326/sdks/java/extensions/google-cloud-platform-core/src/main/java/org/apache/beam/sdk/util/GcsUtil.java#L551>
>>> wrapper,
>>> which leaks memory. The question is, why is that wrapper necessary
>>> if executor.shutdown(); is later unconditionally called?
>>>
>>
