flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Aljoscha Krettek <aljos...@apache.org>
Subject Re: RocksDB Statebackend
Date Wed, 13 Apr 2016 08:16:06 GMT
Hi Maxim,
yes the plan is to have a cache of hot values that uses the managed memory
abstraction of Flink so that we can make sure that we stay within memory
bounds and don't run into OOM exceptions.

On Tue, 12 Apr 2016 at 23:37 Maxim <mfateev@gmail.com> wrote:

> Is it possible to add an option to store the state in the Java HashMap and
> write its content to RocksDB when checkpointing? For "hot" keys that are
> updated very frequently such optimization would help with performance.
>
> I know that you are also working on incremental checkpoints which would
> also be big win for jobs with a large number of keys.
>
> Thanks,
>
> Maxim.
>
> On Tue, Apr 12, 2016 at 10:39 AM, Stephan Ewen <sewen@apache.org> wrote:
>
>> Concerning the size of RocksDB snapshots - I am wondering if RocksDB
>> simply does not compact for a long time, thus having a lot of stale data in
>> the snapshot.
>>
>> That would be especially the case, if you have a lot of changing values
>> for the same set of keys.
>>
>> On Tue, Apr 12, 2016 at 6:41 PM, Aljoscha Krettek <aljoscha@apache.org>
>> wrote:
>>
>>> Hi,
>>> I'm going to try and respond to each point:
>>>
>>> 1. This seems strange, could you give some background on parallelism,
>>> number of operators with state and so on? Also, I'm assuming you are using
>>> the partitioned state abstraction, i.e. getState(), correct?
>>>
>>> 2. your observations are pretty much correct. The reason why RocksDB is
>>> slower is that the FsStateBackend basically stores the state in a Java
>>> HashMap and writes the contents to HDFS when checkpointing. RocksDB stores
>>> data in on-disk files and goes to them for every state access (of course
>>> there are caches, but generally it is like this). I'm actually impressed
>>> that it is still this fast in comparison.
>>>
>>> 3. see 1. (I think for now)
>>>
>>> 4. The checkpointing time is the time from the JobManager deciding to
>>> start a checkpoint until all tasks have confirmed that checkpoint. I have
>>> seen this before and I think it results from back pressure. The problem is
>>> that the checkpoint messages that we sent through the topology are sitting
>>> at the sources because they are also back pressured by the slow processing
>>> of normal records. You should be able to see the actual checkpointing times
>>> (both synchronous and asynchronous) in the log files of the task managers,
>>> they should be very much lower.
>>>
>>> I can go into details, I'm just writing this quickly before calling it a
>>> day. :-)
>>>
>>> Cheers,
>>> Aljoscha
>>>
>>> On Tue, 12 Apr 2016 at 18:21 Konstantin Knauf <
>>> konstantin.knauf@tngtech.com> wrote:
>>>
>>>> Hi everyone,
>>>>
>>>> my experience with RocksDBStatebackend have left me a little bit
>>>> confused. Maybe you guys can confirm that my epxierence is the expected
>>>> behaviour ;):
>>>>
>>>> I have run a "performancetest" twice, once with FsStateBackend and once
>>>> RocksDBStatebackend in comparison. In this particular test the state
>>>> saved is generally not large (in a production scenario it will be
>>>> larger.)
>>>>
>>>> These are my observations:
>>>>
>>>> 1. Minimal Checkpoint Size (no records) with RocksDB was 33MB compared
>>>> to <<1MB with the FSStatebackend.
>>>>
>>>> 2. Throughput dropped from 28k/s -> 18k/s on a small cluster.
>>>>
>>>> 3. Checkpoint sizes as reported in the Dashboard was ca. 1MB for
>>>> FsStatebackend but >100MB for RocksDbStatebackend. I hope the difference
>>>> gets smaller for very large state. Can you confirm?
>>>>
>>>> 4. Checkpointing Times as reported in the Dashboard were 26secs for
>>>> RocksDB during the test and <1 second for FsStatebackend. Does the
>>>> reported time correspond to the sync. + asynchronous part of the
>>>> checkpointing in case of RocksDB? Is there any way to tell how long the
>>>> synchronous part takes?
>>>>
>>>> Form these first observations RocksDB does seem to bring a large
>>>> overhead for state < 1GB, I guess? Is this expected?
>>>>
>>>> Cheers,
>>>>
>>>> Konstantin
>>>>
>>>
>>
>

Mime
View raw message