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

Switch to Threaded View
HBase >> mail # user >> Uneven write request to regions

Copy link to this message
Re: Uneven write request to regions
Re: "The 32 limit makes HBase go into
stress mode, and dump all involving regions contains in those 32 WAL Files."

Pardon, I haven't read all your data points/details thoroughly, but the
above statement is not true. Rather, it looks at the oldest WAL file, and
flushes those regions which would free that WAL file.

But I agree that in general with this kind of workload, we should handle
WAL files more intelligently and free up those WAL files which don't have
any dependency (that is, all their entries are already flushed) when
archiving. We do that in trunk but not in any released version, though.

On Sat, Nov 16, 2013 at 11:16 AM, Asaf Mesika <[EMAIL PROTECTED]> wrote:

> First I forgot to mention that <customerId> in our case is
> MD5(<customerId>).
> In our case, we have so much data flowing in, that we end up having a
> region per <customerId><bucket> pretty quickly and even that, is splitted
> into different regions by specific date duration (timestamp).
> We're not witnessing a hotspot issue. I built some scripts in java and awk,
> and saw that 66% of our customers use more than 1Rs.
> We have two main serious issues: primary and secondary.
> Our primary issue being the slow-region vs fast-region. First let's be
> reminded that a region represents as I detailed before a specific
> <customerId><bucket>. Some customers gets x50 times more data that other
> customers at a specific time frame (2hrs - 1 day). So in a one RS, we have
> regions getting 10 write requests per hour, vs 50k write requests per hour.
> So the region mapped to the slow-filling customer id, doesn't get to the
> 256MB flush limit and hence isn't flushed, while the regions mapped to the
> fast-filling customer id, are flushing very quickly since they are filling
> very quickly.
> Let's say the 1st WAL file contains the put of a slow-filling customerId.
> the fast-filling customerId, fills up the rest of that file. After 20-30
> seconds, the file gets rolled, and another file fills up with fast filling
> customerId. After a while, we get to 32 WAL Files. The 1st file wasn't
> deleted since its region wasn't flushed. The 32 limit makes HBase go into
> stress mode, and dump all involving regions contains in those 32 WAL Files.
> In our case, we saw that it flushes 111 regions. Lots of the store files
> are 3k-3mb sized. So our compaction queue start filling up with those store
> files needs to be compacted.
> At the of the road, the RS gets dead.
> Our secondary issue is those of empty regions - we get to a situation where
> a region is mapped to a specific <customerId>, <bucket>, and date range
> (1/7 - 3/7). Those when we are in August (we TTL set to 30 days), those
> regions gets empty and will never get filled again.
> We assume this somehow wrecks havoc in the load balancer, and also MSLAB
> probably steals 1-2 GB of memory for those empty regions.
> Thanks!
> On Sat, Nov 16, 2013 at 7:25 PM, Mike Axiak <[EMAIL PROTECTED]> wrote:
> > Hi,
> >
> > One new key pattern that we're starting to use is a salt based on a
> shard.
> > For example, let's take your key:
> >
> >   <customerId><bucket><timestampInMs><uniqueId>
> >
> > Consider a shard between 0 and 15 inclusive. We determine this with:
> >
> >  <shard> = abs(hash32(uniqueId) % 16)
> >
> > We can then define a salt to be based on customerId and the shard:
> >
> >  <salt> = hash32(<shard><customerId>)
> >
> > So then the new key becomes:
> >
> >  <salt><customerId><timestampInMs><uniqueId>
> >
> > This will distribute the data for a given customer across the N shards
> that
> > you pick, while having a deterministic function for a given row key (so
> > long as the # of shards you pick is fixed, otherwise you can migrate the
> > data). Placing the bucket after the customerId doesn't help distribute
> the
> > single customer's data at all. Furthermore, by using a separate hash
> > (instead of just <shard><customerId>),  you're guaranteeing that new data
> > will appear in a somewhat random location (i.e., solving the problem of