It fixed the foreman issue perfectly and has significantly increase performance in our test cases. We're still struggling a bit with the data affinity challenge, but it may be unrelated to Drill (as in our environment the name/data nodes are doing the normal HDFS work as well, so it may be a contention issue).
The shuffling is probably not the most optimal way to balance load, but it's better than just picking the first as the foreman each time. On Thu, Apr 16, 2015 at 10:29 AM, Jacques Nadeau <jacq...@apache.org> wrote: > It doesn't currently have plan caching but a simple implementation probably > wouldn't be that difficult (assuming you keep it node-level as opposed to > cluster level). We merged the auto shuffling per session so let us know > how that looks. > > On Wed, Apr 15, 2015 at 4:35 PM, Adam Gilmore <dragoncu...@gmail.com> > wrote: > > > The workload does involve a fair number of short queries. Although when > I > > say short, I'm talking about querying 2-10 million record Parquet files, > so > > they're not extremely short. > > > > Does Drill have plan caching built in at this stage? Might help us > reduce > > some of that foreman overhead. > > > > On Tue, Apr 14, 2015 at 3:02 AM, Jacques Nadeau <jacq...@apache.org> > > wrote: > > > > > Yeah, it seems that way. We should get your patch merged. I just > > reviewed > > > and lgtm. > > > > > > What type of workload are you running? Unless your workload is > planning > > > heavy (e.g. lots of short queries) or does a lot of sorts (the last > merge > > > is on the foreman node), work should be reasonably distributed. > > > > > > On Sun, Apr 12, 2015 at 10:29 PM, Adam Gilmore <dragoncu...@gmail.com> > > > wrote: > > > > > > > Looks like this definitely is the following bug: > > > > > > > > https://issues.apache.org/jira/browse/DRILL-2512 > > > > > > > > It's a pretty severe performance bottleneck having the foreman doing > so > > > > much work. In our environment, the foreman hits basically 95-100% > CPU > > > > while the other drillbits barely do much work. Means it's nearly > > > > impossible for us to scale out. > > > > > > > > On Wed, Apr 8, 2015 at 3:58 PM, Adam Gilmore <dragoncu...@gmail.com> > > > > wrote: > > > > > > > > > Anyone have any more thoughts on this? Anywhere I can start trying > > to > > > > > troubleshoot? > > > > > > > > > > On Thu, Mar 26, 2015 at 4:13 PM, Adam Gilmore < > dragoncu...@gmail.com > > > > > > > > wrote: > > > > > > > > > >> So there are 5 Parquet files, each ~125mb - not sure what I can > > > provide > > > > >> re the block locations? I believe it's under the HDFS block size > so > > > > they > > > > >> should be stored contiguously. > > > > >> > > > > >> I've tried setting the affinity factor to various values (1, 0, > > etc.) > > > > but > > > > >> nothing seems to change that. It always prefers certain nodes. > > > > >> > > > > >> Moreover, we added a stack more nodes and it started picking very > > > > >> specific nodes as foremen (perhaps 2-3 nodes out of 20 were always > > > > picked > > > > >> as foremen). Therefore, the foremen were being swamped with CPU > > while > > > > the > > > > >> other nodes were doing very little work. > > > > >> > > > > >> On Thu, Mar 26, 2015 at 12:12 PM, Steven Phillips < > > > > sphill...@maprtech.com > > > > >> > wrote: > > > > >> > > > > >>> Actually, I believe a query submitted through REST interface will > > > > >>> instantiate a DrillClient, which uses the same > ZKClusterCoordinator > > > > that > > > > >>> sqlline uses, and thus the foreman for the query is not > necessarily > > > on > > > > >>> the > > > > >>> same drillbit as it was submitted to. But I'm still not sure it's > > > > related > > > > >>> to DRILL-2512. > > > > >>> > > > > >>> I'll wait for your additional info before speculating further. > > > > >>> > > > > >>> On Wed, Mar 25, 2015 at 6:54 PM, Adam Gilmore < > > dragoncu...@gmail.com > > > > > > > > >>> wrote: > > > > >>> > > > > >>> > We actually setup a separate load balancer for port 8047 (we're > > > > >>> submitting > > > > >>> > these queries via the REST API at the moment) so Zookeeper etc. > > is > > > > out > > > > >>> of > > > > >>> > the equation, thus I doubt we're hitting DRILL-2512. > > > > >>> > > > > > >>> > When shutitng down the "troublesome" drillbit, it starts > > > > parallelizing > > > > >>> much > > > > >>> > nicer again. We even added 10+ nodes to the cluster and as > long > > as > > > > >>> that > > > > >>> > particular drillbit is shut down, it distributes very nicely. > > The > > > > >>> minute > > > > >>> > we start the drillbit on that node again, it starts swamping it > > > with > > > > >>> work. > > > > >>> > > > > > >>> > I'll shoot through the JSON profiles and some more information > on > > > the > > > > >>> > dataset etc. later today (Australian time!). > > > > >>> > > > > > >>> > On Thu, Mar 26, 2015 at 5:31 AM, Steven Phillips < > > > > >>> sphill...@maprtech.com> > > > > >>> > wrote: > > > > >>> > > > > > >>> > > I didn't notice at first that Adam said "no matter who the > > > foreman > > > > >>> is". > > > > >>> > > > > > > >>> > > Another suspicion I have is that our current logic for > > assigning > > > > work > > > > >>> > will > > > > >>> > > assign to the exact same nodes every time we query a > particular > > > > >>> table. > > > > >>> > > Changing affinity factor may change it, but it will still be > > the > > > > same > > > > >>> > every > > > > >>> > > time. That is my suspicion, but I am not sure why shutting > down > > > the > > > > >>> > > drillbit would improve performance. I would expect that > > shutting > > > > >>> down the > > > > >>> > > drillbit would result in a different drillbit becoming the > > > hotspot. > > > > >>> > > > > > > >>> > > On Wed, Mar 25, 2015 at 12:16 PM, Jacques Nadeau < > > > > jacq...@apache.org > > > > >>> > > > > > >>> > > wrote: > > > > >>> > > > > > > >>> > > > On Steven's point, the node that the client connects to is > > not > > > > >>> > currently > > > > >>> > > > randomized. Given your description of behavior, I'm not > sure > > > > that > > > > >>> > you're > > > > >>> > > > hitting 2512 or just general undesirable distribution. > > > > >>> > > > > > > > >>> > > > On Wed, Mar 25, 2015 at 10:18 AM, Steven Phillips < > > > > >>> > > sphill...@maprtech.com> > > > > >>> > > > wrote: > > > > >>> > > > > > > > >>> > > > > This is a known issue: > > > > >>> > > > > > > > > >>> > > > > https://issues.apache.org/jira/browse/DRILL-2512 > > > > >>> > > > > > > > > >>> > > > > On Wed, Mar 25, 2015 at 8:13 AM, Andries Engelbrecht < > > > > >>> > > > > aengelbre...@maprtech.com> wrote: > > > > >>> > > > > > > > > >>> > > > > > What version of Drill are you running? > > > > >>> > > > > > > > > > >>> > > > > > Any hints when looking at the query profiles? Is the > node > > > > that > > > > >>> is > > > > >>> > > being > > > > >>> > > > > > hammered the foreman for the queries and most of the > > major > > > > >>> > fragments > > > > >>> > > > are > > > > >>> > > > > > tied to the foreman? > > > > >>> > > > > > > > > > >>> > > > > > —Andries > > > > >>> > > > > > > > > > >>> > > > > > > > > > >>> > > > > > On Mar 25, 2015, at 12:00 AM, Adam Gilmore < > > > > >>> dragoncu...@gmail.com> > > > > >>> > > > > wrote: > > > > >>> > > > > > > > > > >>> > > > > > > Hi guys, > > > > >>> > > > > > > > > > > >>> > > > > > > I'm trying to understand how this could be > possible. I > > > > have > > > > >>> a > > > > >>> > > Hadoop > > > > >>> > > > > > > cluster of a name node and two data nodes setup. All > > > have > > > > >>> > > identical > > > > >>> > > > > > specs > > > > >>> > > > > > > in terms of CPU/RAM etc. > > > > >>> > > > > > > > > > > >>> > > > > > > The two data nodes have a replicated HDFS setup where > > I'm > > > > >>> storing > > > > >>> > > > some > > > > >>> > > > > > > Parquet files. > > > > >>> > > > > > > > > > > >>> > > > > > > A Drill cluster (with Zookeeper) is running with > > > Drillbits > > > > >>> on all > > > > >>> > > > three > > > > >>> > > > > > > servers. > > > > >>> > > > > > > > > > > >>> > > > > > > When I submit a query to *any* of the Drillbits, no > > > matter > > > > >>> who > > > > >>> > the > > > > >>> > > > > > foreman > > > > >>> > > > > > > is, one particular data node gets picked to do the > vast > > > > >>> majority > > > > >>> > of > > > > >>> > > > the > > > > >>> > > > > > > work. > > > > >>> > > > > > > > > > > >>> > > > > > > We've even added three more task nodes to the cluster > > and > > > > >>> > > everything > > > > >>> > > > > > still > > > > >>> > > > > > > puts a huge load on one particular server. > > > > >>> > > > > > > > > > > >>> > > > > > > There is nothing unique about this data node. HDFS > is > > > > fully > > > > >>> > > > replicated > > > > >>> > > > > > (no > > > > >>> > > > > > > unreplicated blocks) to the other data node. > > > > >>> > > > > > > > > > > >>> > > > > > > I know that Drill tries to get data locality, so I'm > > > > >>> wondering if > > > > >>> > > > this > > > > >>> > > > > is > > > > >>> > > > > > > the cause, but this essentially swamping this data > node > > > > with > > > > >>> 100% > > > > >>> > > CPU > > > > >>> > > > > > usage > > > > >>> > > > > > > while leaving the others barely doing any work. > > > > >>> > > > > > > > > > > >>> > > > > > > As soon as we shut down the Drillbit on this data > node, > > > > query > > > > >>> > > > > performance > > > > >>> > > > > > > increases significantly. > > > > >>> > > > > > > > > > > >>> > > > > > > Any thoughts on how I can troubleshoot why Drill is > > > picking > > > > >>> that > > > > >>> > > > > > particular > > > > >>> > > > > > > node? > > > > >>> > > > > > > > > > >>> > > > > > > > > > >>> > > > > > > > > >>> > > > > > > > > >>> > > > > -- > > > > >>> > > > > Steven Phillips > > > > >>> > > > > Software Engineer > > > > >>> > > > > > > > > >>> > > > > mapr.com > > > > >>> > > > > > > > > >>> > > > > > > > >>> > > > > > > >>> > > > > > > >>> > > > > > > >>> > > -- > > > > >>> > > Steven Phillips > > > > >>> > > Software Engineer > > > > >>> > > > > > > >>> > > mapr.com > > > > >>> > > > > > > >>> > > > > > >>> > > > > >>> > > > > >>> > > > > >>> -- > > > > >>> Steven Phillips > > > > >>> Software Engineer > > > > >>> > > > > >>> mapr.com > > > > >>> > > > > >> > > > > >> > > > > > > > > > > > > > > >
