kudu-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Chris George" <Christopher.Geo...@rms.com>
Subject Re: Spark on Kudu
Date Mon, 11 Apr 2016 19:21:56 GMT
I have a wip kuduRDD that I made a few months ago. I pushed it into gerrit if you want to take
a look. http://gerrit.cloudera.org:8080/#/c/2754/
It does pushdown predicates which the existing input formatter based rdd does not.

Within the next two weeks I’m planning to implement a datasource for spark that will have
pushdown predicates and insertion/update functionality (need to look more at cassandra and
the hbase datasource for best way to do this) I agree that server side upsert would be helpful.
Having a datasource would give us useful data frames and also make spark sql usable for kudu.

My reasoning for having a spark datasource and not using Impala is: 1. We have had trouble
getting impala to run fast with high concurrency when compared to spark 2. We interact with
datasources which do not integrate with impala. 3. We have custom sql query planners for extended
sql functionality.

-Chris George

On 4/11/16, 12:22 PM, "Jean-Daniel Cryans" <jdcryans@apache.org<mailto:jdcryans@apache.org>>

You guys make a convincing point, although on the upsert side we'll need more support from
the servers. Right now all you can do is an INSERT then, if you get a dup key, do an UPDATE.
I guess we could at least add an API on the client side that would manage it, but it wouldn't
be atomic.


On Mon, Apr 11, 2016 at 9:34 AM, Mark Hamstra <mark@clearstorydata.com<mailto:mark@clearstorydata.com>>
It's pretty simple, actually.  I need to support versioned datasets in a Spark SQL environment.
 Instead of a hack on top of a Parquet data store, I'm hoping (among other reasons) to be
able to use Kudu's write and timestamp-based read operations to support not only appending
data, but also updating existing data, and even some schema migration.  The most typical use
case is a dataset that is updated periodically (e.g., weekly or monthly) in which the the
preliminary data in the previous window (week or month) is updated with values that are expected
to remain unchanged from then on, and a new set of preliminary values for the current window
need to be added/appended.

Using Kudu's Java API and developing additional functionality on top of what Kudu has to offer
isn't too much to ask, but the ease of integration with Spark SQL will gate how quickly we
would move to using Kudu and how seriously we'd look at alternatives before making that decision.

On Mon, Apr 11, 2016 at 8:14 AM, Jean-Daniel Cryans <jdcryans@apache.org<mailto:jdcryans@apache.org>>

Thanks for taking some time to reply in this thread, glad it caught the attention of other

On Sun, Apr 10, 2016 at 12:33 PM, Mark Hamstra <mark@clearstorydata.com<mailto:mark@clearstorydata.com>>
Do they care being able to insert into Kudu with SparkSQL

I care about insert into Kudu with Spark SQL.  I'm currently delaying a refactoring of some
Spark SQL-oriented insert functionality while trying to evaluate what to expect from Kudu.
 Whether Kudu does a good job supporting inserts with Spark SQL will be a key consideration
as to whether we adopt Kudu.

I'd like to know more about why SparkSQL inserts in necessary for you. Is it just that you
currently do it that way into some database or parquet so with minimal refactoring you'd be
able to use Kudu? Would re-writing those SQL lines into Scala and directly use the Java API's
KuduSession be too much work?

Additionally, what do you expect to gain from using Kudu VS your current solution? If it's
not completely clear, I'd love to help you think through it.

On Sun, Apr 10, 2016 at 12:23 PM, Jean-Daniel Cryans <jdcryans@apache.org<mailto:jdcryans@apache.org>>
Yup, starting to get a good idea.

What are your DS folks looking for in terms of functionality related to Spark? A SparkSQL
integration that's as fully featured as Impala's? Do they care being able to insert into Kudu
with SparkSQL or just being able to query real fast? Anything more specific to Spark that
I'm missing?

FWIW the plan is to get to 1.0 in late Summer/early Fall. At Cloudera all our resources are
committed to making things happen in time, and a more fully featured Spark integration isn't
in our plans during that period. I'm really hoping someone in the community will help with
Spark, the same way we got a big contribution for the Flume sink.


On Sun, Apr 10, 2016 at 11:29 AM, Benjamin Kim <bbuild11@gmail.com<mailto:bbuild11@gmail.com>>
Yes, we took Kudu for a test run using 0.6 and 0.7 versions. But, since it’s not “production-ready”,
upper management doesn’t want to fully deploy it yet. They just want to keep an eye on it
though. Kudu was so much simpler and easier to use in every aspect compared to HBase. Impala
was great for the report writers and analysts to experiment with for the short time it was
up. But, once again, the only blocker was the lack of Spark support for our Data Developers/Scientists.
So, production-level data population won’t happen until then.

I hope this helps you get an idea where I am coming from…


On Apr 10, 2016, at 11:08 AM, Jean-Daniel Cryans <jdcryans@apache.org<mailto:jdcryans@apache.org>>

On Sun, Apr 10, 2016 at 12:30 AM, Benjamin Kim <bbuild11@gmail.com<mailto:bbuild11@gmail.com>>

The main thing I hear that Cassandra is being used as an updatable hot data store to ensure
that duplicates are taken care of and idempotency is maintained. Whether data was directly
retrieved from Cassandra for analytics, reports, or searches, it was not clear as to what
was its main use. Some also just used it for a staging area to populate downstream tables
in parquet format. The last thing I heard was that CQL was terrible, so that rules out much
use of direct queries against it.

I'm no C* expert, but I don't think CQL is meant for real analytics, just ease of use instead
of plainly using the APIs. Even then, Kudu should beat it easily on big scans. Same for HBase.
We've done benchmarks against the latter, not the former.

As for our company, we have been looking for an updatable data store for a long time that
can be quickly queried directly either using Spark SQL or Impala or some other SQL engine
and still handle TB or PB of data without performance degradation and many configuration headaches.
For now, we are using HBase to take on this role with Phoenix as a fast way to directly query
the data. I can see Kudu as the best way to fill this gap easily, especially being the closest
thing to other relational databases out there in familiarity for the many SQL analytics people
in our company. The other alternative would be to go with AWS Redshift for the same reasons,
but it would come at a cost, of course. If we went with either solutions, Kudu or Redshift,
it would get rid of the need to extract from HBase to parquet tables or export to PostgreSQL
to support more of the SQL language using by analysts or the reporting software we use..

Ok, the usual then *smile*. Looks like we're not too far off with Kudu. Have you folks tried
Kudu with Impala yet with those use cases?

I hope this helps.

It does, thanks for nice reply.


On Apr 9, 2016, at 2:00 PM, Jean-Daniel Cryans <jdcryans@apache.org<mailto:jdcryans@apache.org>>

Ha first time I'm hearing about SMACK. Inside Cloudera we like to refer to "Impala + Kudu"
as Kimpala, but yeah it's not as sexy. My colleagues who were also there did say that the
hype around Spark isn't dying down.

There's definitely an overlap in the use cases that Cassandra, HBase, and Kudu cater to. I
wouldn't go as far as saying that C* is just an interim solution for the use case you describe.

Nothing significant happened in Kudu over the past month, it's a storage engine so things
move slowly *smile*. I'd love to see more contributions on the Spark front. I know there's
code out there that could be integrated in kudu-spark, it just needs to land in gerrit. I'm
sure folks will happily review it.

Do you have relevant experiences you can share? I'd love to learn more about the use cases
for which you envision using Kudu as a C* replacement.



On Fri, Apr 8, 2016 at 12:45 PM, Benjamin Kim <bbuild11@gmail.com<mailto:bbuild11@gmail.com>>
Hi J-D,

My colleagues recently came back from Strata in San Jose. They told me that everything was
about Spark and there is a big buzz about the SMACK stack (Spark, Mesos, Akka, Cassandra,
Kafka). I still think that Cassandra is just an interim solution as a low-latency, easily
queried data store. I was wondering if anything significant happened in regards to Kudu, especially
on the Spark front. Plus, can you come up with your own proposed stack acronym to promote?


On Mar 1, 2016, at 12:20 PM, Jean-Daniel Cryans <jdcryans@apache.org<mailto:jdcryans@apache.org>>

Hi Ben,

AFAIK no one in the dev community committed to any timeline. I know of one person on the Kudu
Slack who's working on a better RDD, but that's about it.



On Tue, Mar 1, 2016 at 11:00 AM, Benjamin Kim <bkim@amobee.com<mailto:bkim@amobee.com>>
Hi J-D,

Quick question… Is there an ETA for KUDU-1214? I want to target a version of Kudu to begin
real testing of Spark against it for our devs. At least, I can tell them what timeframe to

Just curious,
Benjamin Kim
Data Solutions Architect

[a•mo•bee] (n.) the company defining digital marketing.

Mobile: +1 818 635 2900<tel:%2B1%20818%20635%202900>
3250 Ocean Park Blvd, Suite 200  |  Santa Monica, CA 90405  |  www.amobee.com<http://www.amobee.com/>

On Feb 24, 2016, at 3:51 PM, Jean-Daniel Cryans <jdcryans@apache.org<mailto:jdcryans@apache.org>>

The DStream stuff isn't there at all. I'm not sure if it's needed either.

The kuduRDD is just leveraging the MR input format, ideally we'd use scans directly.

The SparkSQL stuff is there but it doesn't do any sort of pushdown. It's really basic.

The goal was to provide something for others to contribute to. We have some basic unit tests
that others can easily extend. None of us on the team are Spark experts, but we'd be really
happy to assist one improve the kudu-spark code.


On Wed, Feb 24, 2016 at 3:41 PM, Benjamin Kim <bbuild11@gmail.com<mailto:bbuild11@gmail.com>>

It looks like it fulfills most of the basic requirements (kudu RDD, kudu DStream) in KUDU-1214.
Am I right? Besides shoring up more Spark SQL functionality (Dataframes) and doing the documentation,
what more needs to be done? Optimizations?

I believe that it’s a good place to start using Spark with Kudu and compare it to HBase
with Spark (not clean).


On Feb 24, 2016, at 3:10 PM, Jean-Daniel Cryans <jdcryans@apache.org<mailto:jdcryans@apache.org>>

AFAIK no one is working on it, but we did manage to get this in for 0.7.0: https://issues.cloudera.org/browse/KUDU-1321

It's a really simple wrapper, and yes you can use SparkSQL on Kudu, but it will require a
lot more work to make it fast/useful.

Hope this helps,


On Wed, Feb 24, 2016 at 3:08 PM, Benjamin Kim <bbuild11@gmail.com<mailto:bbuild11@gmail.com>>
I see this KUDU-1214<https://issues.cloudera.org/browse/KUDU-1214> targeted for 0.8.0,
but I see no progress on it. When this is complete, will this mean that Spark will be able
to work with Kudu both programmatically and as a client via Spark SQL? Or is there more work
that needs to be done on the Spark side for it to work?

Just curious.


View raw message