Cool, thanks for the report, Ben. For what it's worth, I think there's still some low hanging fruit in the Spark connector for Kudu (for example, I believe locality on reads is currently broken). So, you can expect performance to continue to improve in future versions. I'd also be interested to see results on Kudu for a much larger dataset - my guess is a lot of the 6 seconds you're seeing is constant overhead from Spark job setup, etc, given that the performance doesn't seem to get slower as you went from 700K rows to 13M rows.

-Todd

On Tue, Jun 28, 2016 at 3:03 PM, Benjamin Kim <bbuild11@gmail.com> wrote:
FYI.

I did a quick-n-dirty performance test.

First, the setup:
QA cluster:
  • 15 data nodes
    • 64GB memory each
    • HBase is using 4GB of memory
    • Kudu is using 1GB of memory
  • 1 HBase/Kudu master node
    • 64GB memory
    • HBase/Kudu master is using 1GB of memory each
  • 10Gb Ethernet

Using Spark on both to load/read events data (84 columns per row), I was able to record performance for each. On the HBase side, I used the Phoenix 4.7 Spark plugin where DataFrames can be used directly. On the Kudu side, I used the Spark connector. I created an events table in Phoenix using the CREATE TABLE statement and created the equivalent in Kudu using the Spark method based off of a DataFrame schema.

Here are the numbers for Phoenix/HBase.
1st run:
> 715k rows
- write: 2.7m

> 715k rows in HBase table
- read: 0.1s
- count: 3.8s
- aggregate: 61s

2nd run:
> 5.2M rows
- write: 11m
* had 4 region servers go down, had to retry the 5.2M row write

> 5.9M rows in HBase table
- read: 8s
- count: 3m
- aggregate: 46s

3rd run:
> 6.8M rows
- write: 9.6m

> 12.7M rows
- read: 10s
- count: 3m
- aggregate: 44s


Here are the numbers for Kudu.
1st run:
> 715k rows
- write: 18s

> 715k rows in Kudu table
- read: 0.2s
- count: 18s
- aggregate: 5s

2nd run:
> 5.2M rows
- write: 33s

> 5.9M rows in Kudu table
- read: 0.2s
- count: 16s
- aggregate: 6s

3rd run:
> 6.8M rows
- write: 27s

> 12.7M rows in Kudu table
- read: 0.2s
- count: 16s
- aggregate: 6s

The Kudu results are impressive if you take these number as-is. Kudu is close to 18x faster at writing (UPSERT). Kudu is 30x faster at reading (HBase times increase as data size grows).  Kudu is 7x faster at full row counts. Lastly, Kudu is 3x faster doing an aggregate query (count distinct event_id’s per user_id). *Remember that this is small cluster, times are still respectable for both systems, HBase could have been configured better, and the HBase table could have been better tuned.

Cheers,
Ben


On Jun 15, 2016, at 10:13 AM, Dan Burkert <dan@cloudera.com> wrote:

Adding partition splits when range partitioning is done via the CreateTableOptions.addSplitRow method.  You can find more about the different partitioning options in the schema design guide.  We generally recommend sticking to hash partitioning if possible, since you don't have to determine your own split rows.

- Dan

On Wed, Jun 15, 2016 at 9:17 AM, Benjamin Kim <bbuild11@gmail.com> wrote:
Todd,

I think the locality is not within our setup. We have the compute cluster with Spark, YARN, etc. on its own, and we have the storage cluster with HBase, Kudu, etc. on another. We beefed up the hardware specs on the compute cluster and beefed up storage capacity on the storage cluster. We got this setup idea from the Databricks folks. I do have a question. I created the table to use range partition on columns. I see that if I use hash partition I can set the number of splits, but how do I do that using range (50 nodes * 10 = 500 splits)?

Thanks,
Ben


On Jun 15, 2016, at 9:11 AM, Todd Lipcon <todd@cloudera.com> wrote:

Awesome use case. One thing to keep in mind is that spark parallelism will be limited by the number of tablets. So, you might want to split into 10 or so buckets per node to get the best query throughput.

Usually if you run top on some machines while running the query you can see if it is fully utilizing the cores.

Another known issue right now is that spark locality isn't working properly on replicated tables so you will use a lot of network traffic. For a perf test you might want to try a table with replication count 1

On Jun 15, 2016 5:26 PM, "Benjamin Kim" <bbuild11@gmail.com> wrote:
Hi Todd,

I did a simple test of our ad events. We stream using Spark Streaming directly into HBase, and the Data Analysts/Scientists do some insight/discovery work plus some reports generation. For the reports, we use SQL, and the more deeper stuff, we use Spark. In Spark, our main data currency store of choice is DataFrames.

The schema is around 83 columns wide where most are of the string data type.

"event_type", "timestamp", "event_valid", "event_subtype", "user_ip", "user_id", "mappable_id",
"cookie_status", "profile_status", "user_status", "previous_timestamp", "user_agent", "referer",
"host_domain", "uri", "request_elapsed", "browser_languages", "acamp_id", "creative_id",
"location_id", “pcamp_id",
"pdomain_id", "continent_code", "country", "region", "dma", "city", "zip", "isp", "line_speed",
"gender", "year_of_birth", "behaviors_read", "behaviors_written", "key_value_pairs", "acamp_candidates",
"tag_format", "optimizer_name", "optimizer_version", "optimizer_ip", "pixel_id", “video_id",
"video_network_id", "video_time_watched", "video_percentage_watched", "video_media_type",
"video_player_iframed", "video_player_in_view", "video_player_width", "video_player_height",
"conversion_valid_sale", "conversion_sale_amount", "conversion_commission_amount", "conversion_step",
"conversion_currency", "conversion_attribution", "conversion_offer_id", "custom_info", "frequency",
"recency_seconds", "cost", "revenue", “optimizer_acamp_id",
"optimizer_creative_id", "optimizer_ecpm", "impression_id", "diagnostic_data",
"user_profile_mapping_source", "latitude", "longitude", "area_code", "gmt_offset", "in_dst",
"proxy_type", "mobile_carrier", "pop", "hostname", "profile_expires", "timestamp_iso", "reference_id",
"identity_organization", "identity_method"

Most queries are like counts of how many users use what browser, how many are unique users, etc. The part that scares most users is when it comes to joining this data with other dimension/3rd party events tables because of shear size of it.

We do what most companies do, similar to what I saw in earlier presentations of Kudu. We dump data out of HBase into partitioned Parquet tables to make query performance manageable.

I will coordinate with a data scientist today to do some tests. He is working on identity matching/record linking of users from 2 domains: US and Singapore, using probabilistic deduping algorithms. I will load the data from ad events from both countries, and let him run his process against this data in Kudu. I hope this will “wow” the team.

Thanks,
Ben

On Jun 15, 2016, at 12:47 AM, Todd Lipcon <todd@cloudera.com> wrote:

Hi Benjamin,

What workload are you using for benchmarks? Using spark or something more custom? rdd or data frame or SQL, etc? Maybe you can share the schema and some queries

Todd

Todd

On Jun 15, 2016 8:10 AM, "Benjamin Kim" <bbuild11@gmail.com> wrote:
Hi Todd,

Now that Kudu 0.9.0 is out. I have done some tests. Already, I am impressed. Compared to HBase, read and write performance are better. Write performance has the greatest improvement (> 4x), while read is > 1.5x. Albeit, these are only preliminary tests. Do you know of a way to really do some conclusive tests? I want to see if I can match your results on my 50 node cluster.

Thanks,
Ben

On May 30, 2016, at 10:33 AM, Todd Lipcon <todd@cloudera.com> wrote:

On Sat, May 28, 2016 at 7:12 AM, Benjamin Kim <bbuild11@gmail.com> wrote:
Todd,

It sounds like Kudu can possibly top or match those numbers put out by Aerospike. Do you have any performance statistics published or any instructions as to measure them myself as good way to test? In addition, this will be a test using Spark, so should I wait for Kudu version 0.9.0 where support will be built in?

We don't have a lot of benchmarks published yet, especially on the write side. I've found that thorough cross-system benchmarks are very difficult to do fairly and accurately, and often times users end up misguided if they pay too much attention to them :) So, given a finite number of developers working on Kudu, I think we've tended to spend more time on the project itself and less time focusing on "competition". I'm sure there are use cases where Kudu will beat out Aerospike, and probably use cases where Aerospike will beat Kudu as well.

From my perspective, it would be great if you can share some details of your workload, especially if there are some areas you're finding Kudu lacking. Maybe we can spot some easy code changes we could make to improve performance, or suggest a tuning variable you could change.

-Todd


On May 27, 2016, at 9:19 PM, Todd Lipcon <todd@cloudera.com> wrote:

On Fri, May 27, 2016 at 8:20 PM, Benjamin Kim <bbuild11@gmail.com> wrote:
Hi Mike,

First of all, thanks for the link. It looks like an interesting read. I checked that Aerospike is currently at version 3.8.2.3, and in the article, they are evaluating version 3.5.4. The main thing that impressed me was their claim that they can beat Cassandra and HBase by 8x for writing and 25x for reading. Their big claim to fame is that Aerospike can write 1M records per second with only 50 nodes. I wanted to see if this is real.

1M records per second on 50 nodes is pretty doable by Kudu as well, depending on the size of your records and the insertion order. I've been playing with a ~70 node cluster recently and seen 1M+ writes/second sustained, and bursting above 4M. These are 1KB rows with 11 columns, and with pretty old HDD-only nodes. I think newer flash-based nodes could do better.
 

To answer your questions, we have a DMP with user profiles with many attributes. We create segmentation information off of these attributes to classify them. Then, we can target advertising appropriately for our sales department. Much of the data processing is for applying models on all or if not most of every profile’s attributes to find similarities (nearest neighbor/clustering) over a large number of rows when batch processing or a small subset of rows for quick online scoring. So, our use case is a typical advanced analytics scenario. We have tried HBase, but it doesn’t work well for these types of analytics.

I read, that Aerospike in the release notes, they did do many improvements for batch and scan operations.

I wonder what your thoughts are for using Kudu for this.

Sounds like a good Kudu use case to me. I've heard great things about Aerospike for the low latency random access portion, but I've also heard that it's _very_ expensive, and not particularly suited to the columnar scan workload. Lastly, I think the Apache license of Kudu is much more appealing than the AGPL3 used by Aerospike. But, that's not really a direct answer to the performance question :)
 

Thanks,
Ben


On May 27, 2016, at 6:21 PM, Mike Percy <mpercy@cloudera.com> wrote:

Have you considered whether you have a scan heavy or a random access heavy workload? Have you considered whether you always access / update a whole row vs only a partial row? Kudu is a column store so has some awesome performance characteristics when you are doing a lot of scanning of just a couple of columns.

I don't know the answer to your question but if your concern is performance then I would be interested in seeing comparisons from a perf perspective on certain workloads.

Finally, a year ago Aerospike did quite poorly in a Jepsen test: https://aphyr.com/posts/324-jepsen-aerospike

I wonder if they have addressed any of those issues.

Mike

On Friday, May 27, 2016, Benjamin Kim <bbuild11@gmail.com> wrote:
I am just curious. How will Kudu compare with Aerospike (http://www.aerospike.com)? I went to a Spark Roadshow and found out about this piece of software. It appears to fit our use case perfectly since we are an ad-tech company trying to leverage our user profiles data. Plus, it already has a Spark connector and has a SQL-like client. The tables can be accessed using Spark SQL DataFrames and, also, made into SQL tables for direct use with Spark SQL ODBC/JDBC Thriftserver. I see from the work done here http://gerrit.cloudera.org:8080/#/c/2992/ that the Spark integration is well underway and, from the looks of it lately, almost complete. I would prefer to use Kudu since we are already a Cloudera shop, and Kudu is easy to deploy and configure using Cloudera Manager. I also hope that some of Aerospike’s speed optimization techniques can make it into Kudu in the future, if they have not been already thought of or included.

Just some thoughts…

Cheers,
Ben


-- 
--
Mike Percy
Software Engineer, Cloudera






-- 
Todd Lipcon
Software Engineer, Cloudera




-- 
Todd Lipcon
Software Engineer, Cloudera








--
Todd Lipcon
Software Engineer, Cloudera