drill-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Benjamin Schaff <benjamin.sch...@reactivecore.com>
Subject Re: Embedding Drill as a distributed query engine
Date Wed, 22 Jan 2020 14:28:36 GMT
Hi, thanks everyone for the feedback.

The current database query API support pushdowns (filtering and
projections) but when dealing with billion rows, it's still a lot to move
over the network.
The RPC API itself is not the performance bottleneck, we have our own
binary format similar to flatbuffer with query time codegen readers and
writers so that part is ok.

On the question about why the spark part is kind of slow, I do batch
(usually around 50k rows at a time) but my guess is that going from our
binary format to spark internal row format and then spark moves it to
unsaferow is a lot of transformation for "nothing".
We have a codegen parser that does internal format to spark row format but
directly speaking unsaferow is much more involved so I put it on the side
for now.

Here is what I am going to try from all the feedback you gave me:
1) Since premature optimization is the root of evil, and my spark
assumption might not hold true for Drill, I will try to do a "remote"
2) I will try to see if I can use Drill internal format to ship it on the
network, if anybody could be kind enough to give me a pointer where to look
that would be awesome
3) I will upgrade my current integration to merge the "remote" with the
"local" one

I will keep you guys updated and publish my results so that I can give back
some of my experiments.

On a separate note, I was wondering if/how it was possible for Drill
(probably hacking somewhere in calcite plan, to push down the joins filter
parts or if it is done automatically)

Again, any idea or comment is welcome.


On Wed, Jan 22, 2020 at 1:28 AM Ted Dunning <ted.dunning@gmail.com> wrote:

> Hmmm....
> I disagree with a lot of what Paul says.
> Here is where I agree fully:
> 1) collocating processes in the same JVM increases the blast radius of
> failures. If either the DB or the Drill threads go south, it will take the
> other out. This is a relatively low probability event, but increasing the
> probability, or, worse, coupling the probabilities isn't necessary. On a
> very closely related note, the blast radius of GC is also coupled between
> the two processes.
> 2) lack of control over either process or memory for either process will
> affect the other. That would be bad. See (1).
> 3) coupled scaling is sub-optimal. But that might be compensated for by the
> close coupling of within process communication.
> Where I disagree is how serious these considerations are. Drill is fairly
> well disciplined in terms of heap and off-heap space. Presumably the DB is
> as well. That would mean that the likely impact of (2) would be very small.
> The ease of communication between threads within the same process is
> dramatically better than communication between processes, even
> (especially?) with shared memory.
> My own recommendation would be to *allow* collocation but not assume it.
> Allow for non-collocated Drill bits as well. That allows you to pivot at
> any point.
> On the other hand
> On Tue, Jan 21, 2020 at 5:10 PM Paul Rogers <par0328@yahoo.com.invalid>
> wrote:
> > Hi Benjamin,
> >
> > Very cool project! Drill works well on top of custom data sources.
> >
> > That said, I suspect that actually running Drill inside your process will
> > lead to a large amount of complexity. Your comment focuses on code
> issues.
> > However, there are larger concerns. Although we think of Drill as a
> simple
> > single-threaded, single node tool (when run in SqlLine or on a Mac),
> Drill
> > is designed to be fully distributed.
> >
> > As queries get larger, you will find that Drill itself uses large amounts
> > of memory and CPU to run a query quickly. (Imagine a join or sort of
> > billions of rows from several tables.) Drill has its own memory
> management
> > system to handle the large blocks of memory needed. Your DB also needs
> > memory. You'd need a way to unify Drill's memory management with your own
> > -- a daunting task.
> >
> > Grinding through billions of rows is CPU intensive. Drill manages its own
> > thread and makes very liberal use of CPU. Your DB engine likely also has
> a
> > threading model. Again, integrating the two is difficult. We could go on.
> >
> > In short, although Drill works well as a query engine on top of a custom
> > data source; Drill itself is not designed to be a library included into
> > your app process; it is designed to run as its own distributed set of
> > processes running alongside your process.
> >
> > We could, of course, change the design, but that would be a bit of a big
> > project because of the above issues. Might be interesting to think how
> > you'd embed a distributed framework as a library in some host process.
> Not
> > sure I've ever seen this done for any tool. (If anyone knows of an
> example,
> > please let us know.)
> >
> >
> > I wonder if there is a better solution. Run Drill alongside your DB on
> the
> > same nodes. Have Drill then obtain data from your DB via an API. The
> quick
> > & dirty solution is to use an RPC API. You can get fancy and use shared
> > memory. A side benefit is that other tools can also use the API. For
> > example, if you find you need Spark integration, it is easier to provide.
> > (You can't, of course, run Spark in your DB process.)
> >
> > In this case, an "embedded solution" means that Drill is embedded in your
> > app cluster (like ZK), not that it is embedded in your app process.
> >
> >
> > In this way, you can tune Drill's memory and CPU usage separately from
> > that of your engine, making the problem tractable. This model is, in
> fact,
> > very similar to the traditional HDFS model in which both Drill and HDFS
> run
> > on the same nodes. It is also similar to what MapR did with the MapR DB
> > integration.
> >
> >
> > Further, by separating the two, you can run Drill on its own nodes if you
> > find your queries are getting larger and more expensive. That is, you can
> > scale out be separating compute (Drill) from storage (your DB), allowing
> > each to scale independently.
> >
> >
> > And, of course, a failure in one engine (Drill or DB) won't take down the
> > other if the two are in separate processes.
> >
> >
> > In either case, your storage plugin needs to compute data locality. If
> > your DB is distributed, then perhaps it has some scheme for distributing
> > data: hash partitioning, range partitioning, or whatever. Somehow, if I
> > have key 'x', I know to go to node Y to get that value. For example, in
> > HDFS, Drill can distribute block scans to the node(s) with the blocks.
> >
> >
> > Or, maybe data is randomly distributed, so that every scan must run
> > against every DB node; in which case if you have N nodes, you'll run N
> > scans and each will find whatever it happens to contain.
> >
> >
> > If your DB has N nodes, then you need to distribute work to those nodes
> by
> > telling Drill that the max parallelization (reported by the group scan)
> is
> > N. Then, Drill will ask you for the SubScan for each of the N scans, and
> > you can allocate work to those nodes. Either by subsetting the scan (as
> in
> > HDFS) or just running the same scan everywhere.
> >
> >
> > If you go with the two-process model, then your storage plugin can use
> > soft affinity: run the scan on the node that has your DB, else run it on
> > any node and use an RPC to obtain the data. This is how Drill works if it
> > runs on a subset of HDFS nodes.
> >
> > You also asked about the Foreman. At present, Drill assumes nodes are
> > homogeneous: all nodes evenly share work, including the work of the
> > Foreman. Impala, for example, has added a feature to dedicate some nodes
> to
> > be only coordinators (the equivalent of Drill's Foreman). Drill does not
> > yet have that feature.
> >
> > Without the homogeneity assumption, Drill would need some kind of work
> > scheduler to know to give less work to the Forman + Drillbit node and
> more
> > work to the Drillbit-only nodes. Having Foreman-only nodes would keep
> > things simpler. In your ase, such a Foreman would have to reside on a
> node
> > other than one of your DB nodes to keep the DB nodes symmetrical.
> >
> >
> > The above is a high-level survey of the challenges. We'd be happy to
> > discuss specific issues as you refine your design.
> >
> >
> > Thanks,
> > - Paul
> >
> >
> >
> >     On Tuesday, January 21, 2020, 3:00:21 PM PST, Benjamin Schaff <
> > benjamin.schaff@reactivecore.com> wrote:
> >
> >  Hi everyone,
> >
> > I would like to see if you could provide some recommendations/help around
> > integrating Apache Drill as a distributed sql engine in a custom
> database.
> > Maybe I am going about it the wrong way so any feedback is appreciated.
> >
> > What I would like to achieve, is to be able to embed drillbits into my
> > database node, it's a distributed database written mostly in scala so
> it's
> > running inside the jvm. As you would expect, each storage node holds a
> > partition of the data and I would like for each SubScan to be routed to
> the
> > drillbit instance embedded within the database node.
> >
> > At this point, drillbits are running communicating properly with zk (I am
> > using zookeeper for the database also). I can connect to the Plugin I
> > created using sqlline and I can list schemas and tables. So basically,
> all
> > the metadata part is done and working.
> >
> > I managed to build-up the patitionwork and affinity using the distributed
> > metadata off the database and I am stuck in the following situation.
> >
> > If I override the "DistributionAffinity getDistributionAffinity()" method
> > to put it to "HARD", then I end up with having the following error:
> > "IllegalArgumentException: Sender fragment endpoint list should not be
> > empty", and the "applyAssignments" method of the GroupScan receives and
> > empty list of endpoints.
> >
> > If I don't override it then node without "local access" get some work
> > scheduled.
> >
> > I was wondering if there was a way to exclude drillbits to become a
> > foreman.
> >
> > Thanks in advance for any guidance.
> >
> > --
> > *This e-mail and any
> > attachments may contain confidential information and
> > is intended for use solely
> > by the addressee(s).  If you are not the
> >
> > intended recipient of this e-mail, please be aware that any
> dissemination,
> >
> > distribution, copying, or other use of the e-mail in whole or in part, is
> >
> > strictly prohibited.  If you have
> > received this e-mail in error, please
> > notify the sender and permanently delete
> > the original and all copies of the
> > e-mail, attachments, and any printouts. * **

*This e-mail and any
attachments may contain confidential information and 
is intended for use solely
by the addressee(s).  If you are not the

intended recipient of this e-mail, please be aware that any dissemination,

distribution, copying, or other use of the e-mail in whole or in part, is

strictly prohibited.  If you have
received this e-mail in error, please 
notify the sender and permanently delete
the original and all copies of the 
e-mail, attachments, and any printouts. * **

  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message