Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Threaded View
Flume, mail # dev - Review Request: FLUME-1660 Close "idle" hdfs handles

Copy link to this message
Re: Review Request: FLUME-1660 Close "idle" hdfs handles
Juhani Connolly 2012-10-31, 07:46

> On Oct. 29, 2012, 11:49 p.m., Mike Percy wrote:
> > How about we just use the existing timedRollerPool inside the BucketWriter to do this? Just pass closeIdleTimeout to the BucketWriter constructor. At the end of each append, we can just do something like:
> >
> > if (idleCloseFuture != null) idleCloseFuture.cancel(false);
> > idleCloseFuture = timedRollerPool.schedule(new Runnable() {
> >   public void run() {
> >     try {
> >       close();
> >     } catch(Throwable t) {
> >       LOG.error("Unexpected error", t);
> >       if (t instanceof Error) {
> >         throw (Error) t;
> >       }
> >     }
> > }, idleTimeout, TimeUnit.SECONDS);
> >
> > This is basically exactly how the rollInterval timer works (see implementation in BucketWriter.doOpen()). Note you would also want to cancel this future in doClose(), as we do for the rollInterval timer.
> >
> > This approach is certainly slower than just doing a System.getCurrentTimeMillis(), but it's not too bad... Executing future.cancel(false) and future.schedule() seem to take a combined 1.5 microseconds on my laptop. We could put this logic in the doFlush() method and effectively only reset the idle timer at the end of a transaction, which would amortize the cost to almost nil in most cases.
> >
> > The benefit is that if files are rolling too fast, we have a configurable thread pool there to avoid jobs stacking up, whereas a single thread can fall behind. Also, it avoids a synchronization block and iterating through the sfWriters map, and keeps the rolling logic mostly contained in the BucketWriter. It also avoids creating new threads / thread pools.
> Mike Percy wrote:
>     Edit: above should say, at the end of each doFlush() then cancel/reschedule the idleCloseFuture
> Juhani Connolly wrote:
>     Hmm... I can see that as a viable approach but am curious about what happens with the sfWriters map in HDFSEventSink... It seems like old writers are just abandoned there forever? I would like to clean them up properly(I believe this is common in the use case where files are dumped in a file named by date). While not major, this does seem like it would lead to a buildup of inactive writers? We've had OOM errors when running flume with an HDFS sink using the default memory settings. I have no idea if this is related, perhaps it could be? Looks to me that nowhere other than the stop method is the sfWriters map ever cleaned up.
> Juhani Connolly wrote:
>     So, I took a heap dump and checked the retained size for BucketWriter objects... Around 4000 bytes  all told.
>     After a week, one of our aggregator/hdfs sink nodes has got 1500 bucket writers alive in memory, for about 6mb of memory on what are essentially dead objects. This is because we generate a new path(based on time) every hour, for each host/data type. We're still running in a test phase, with only a handful of our servers feeding data, so with more servers and more time, this moderate amount of memory to do nothing.
>     So in the end of the day, at some point, HDFSEventSink does need to get involved to clean this stuff up.
> Juhani Connolly wrote:
>     Uh, that is, 4000 bytes each. Most of that is in the writer.
> Mike Percy wrote:
>     Yeah, you're right about the sfWriters map. The original implementation contained that thing and I never tried to address that issue... it won't cause correctness problems (since close() is effectively idempotent) but yes it will consume some memory. That problem exists with all of the existing rolling code, so it would not just exist with the new code for the close-on-idle feature.
>     One easy band-aid would be to redefine the default maxOpenFiles to, say, 500. That would reduce the severity of the memory reclamation delay, at the limit. If each object takes 4K then a cache of 500 would only take up 2MB which isn't terrible. Another simple approach, which would be a bit ugly (need to be careful with the circular reference) would be to provide a sfWriters reference to each BucketWriter instance, and when the BucketWriter's close() method is called then it can remove itself from the cache if it's still there. Speaking of which, I would prefer to use the Guava CacheBuilder over what we are using now if we can.

How about just running with my original implementation which actually removes inactive writers from the map? It adds another thread but the synchronization block should not be a big deal, as only live writers are iterated over. Considering the polling frequency I don't see this as becoming a problem ever. Regardless of what other rolling may be going on, this should clean up all writers when activated.

As I said, we already have 6mb of memory gone after a week feeding in data from 4 data collection servers to the servers dumping the data to hdfs. We may have up to 100 servers per "cluster"(round robin avro sinked destinations), so that's 150mb of memory leakage per week. It's not really a minor detail when deployed at scale.

I do not think that a default maxOpenFiles is a good idea, simply for the fact that it closes files in order of age, regardless of if they are active or not.
- Juhani
This is an automatically generated e-mail. To reply, visit:
On Oct. 19, 2012, 6:01 a.m., Juhani Connolly wrote: