Speed Issues When Using Driver 2.1.6+

[jwcarman]

First of all, thanks for all your hard work on this library! This really saves us a ton of time!

I'm working on a library which uses cassandra-unit for testing and we're trying to upgrade the Cassandra driver version we're using. From what I can tell, it seems that there is a very noticeable speed degradation when you go from driver version 2.1.5 to 2.1.6 and above. I even start to see test failures ("All host(s) tried for query failed") near the end of the test run. Did you run into this? I'm using these versions:

<cassandra.version>2.1.9</cassandra.version>
<cassandra.driver.version>2.1.6</cassandra.driver.version>
<cassandra.unit.version>2.1.9.2</cassandra.unit.version>

[ilinas]

My experience is that driver 2.1.7 performs much better than 2.1.8 and 2.1.9. It seems that CQL data loading is much slower with the new driver, and especially schema update queries. I am not observing any slowness in production though, just unit testing.

In numbers, my unit tests run in: 40 sec with 2.1.7 2 min 15 sec with 2.1.9

[OrangeDog]

Some of my results for timing cleanEmbeddedCassandra.

• 2.1.9: 4500 ms
• 2.1.7: 2000 ms

Other versions are not compatible with my spring-data-cassandra version, so I cannot test easily.

Profiling with VisualVM suggests the time is spent in Cassandra itself rather than the driver.

org.cassandraunit.utils.EmbeddedCassandraServerHelper.cleanEmbeddedCassandra() 0.0  0.000 ms (0%)   0.000 ms    4,696 ms    0.000 ms
io.netty.util.HashedWheelTimer$Worker.waitForNextTick() 3.7867706   4,696 ms (3.8%) 0.000 ms    4,696 ms    0.000 ms
org.apache.cassandra.utils.concurrent.Transactional$AbstractTransactional.finish()  0.0 0.000 ms (0%)   0.000 ms    4,408 ms    4,408 ms

[olim7t]

@lzvtrifork slow schema updates could be caused by event debouncing (introduced by JAVA-657): before refreshing the driver schema metadata, we wait for a while (1s by default) in case concurrent updates happen, so that we can coalesce them in a single refresh query.

But the downside is that each DDL query now takes at least 1s, not the best for a unit test where a single client does a lot of them. You can try to disable debouncing this way:

Cluster.builder().withQueryOptions(
    new QueryOptions().setRefreshSchemaIntervalMillis(0))

[OrangeDog]

@olim7t that seems to have improved things.

Cassandra-unit should therefore make sure to set that for all DDL operations, such as dropKeyspaces.

Setting the node and nodelist intervals also speeds up connections.

[loucasa]

I'm running into this too, I tried setting the refresh intervals mentioned above and it does make a little difference running a single test class but is still much slower than using 2.1.6

[OrangeDog]

Setting the node and nodelist intervals also speeds up connections.

@jsevellec any reason you didn't?

[jsevellec]

I missed it and didn't get what you mean. Could you be more precise?

[OrangeDog]

@jsevellec these also make it go faster:

.setRefreshNodeListIntervalMillis(0)
.setRefreshNodeIntervalMillis(0)

[jwcarman]

Thanks @jsevellec, I'll give it a spin!