SPARKC-503: Retry PoolBusy Exceptions, Throttle JWCT Calls #1137
Conversation
A change in the Cassandra Java driver has caused the executeAsync method to become fully asynchronous. This means that executeAsync calls can sometimes return an exception of BusyPoolException, meaning that queries were issued and no open execution pathway could be found. These exceptions are always retryable because they indicate the local driver state is temporarly busy. So we will retry these exceptions. To make sure we don't have too many retries we will also make sure to limit the number of max concurrent requests to the current "requestsPerConnection" as specified in the pooling options. We are already setting the maxPool size to the number of cores per executor so this should prohibit having more requests in flight than we have open connections.
|
Build against Scala 2.11 finished with failure |
|
Build against Scala 2.10 finished with success |
|
Build against Scala 2.11 finished with success |
| @@ -194,6 +194,15 @@ class TableWriterSpec extends SparkCassandraITFlatSpecBase { | |||
| verifyKeyValueTable("key_value") | |||
| } | |||
|
|
|||
| it should "pass the exception back to the driver on a write failure" in { | |||
There was a problem hiding this comment.
travis does not like this test https://travis-ci.org/datastax/spark-cassandra-connector/jobs/264511105#L4452
| @@ -18,6 +19,7 @@ class AsyncExecutor[T, R](asyncAction: T => ListenableFuture[R], maxConcurrentTa | |||
|
|
|||
| private val semaphore = new Semaphore(maxConcurrentTasks) | |||
| private val pendingFutures = new TrieMap[ListenableFuture[R], Boolean] | |||
| var latestExeception: Option[Throwable] = None | |||
There was a problem hiding this comment.
could we make this private and expose exc info as read only thing?
| if (_successful) _successful = false | ||
| release() | ||
| settable.setException(throwable) | ||
| latestExeception = Some(throwable) |
There was a problem hiding this comment.
Since _successful volatile write is conditional, we could make _successful not volatile and make latestException volatile. This may make Travis happy :)
There was a problem hiding this comment.
Also it looks like it should be placed in critical section.
There was a problem hiding this comment.
i'm not sure what you mean in "the critical section"
There was a problem hiding this comment.
I'm meant putting it before release(), it's all good now
| throw new IOException( | ||
| s"""Failed to write statements to $keyspaceName.$tableName. The | ||
| |latest exception was | ||
| | ${queryExecutor.latestExeception.map( _.getMessage).getOrElse("Unable to determine")} |
There was a problem hiding this comment.
It looks like _successful (and successful) is redundant. The if could be changed to latestException.map(...)
|
@RussellSpitzer could you also elaborate a bit about
If I understand correctly, max concurrent requests while saving to C* is limited by |
|
Build against Scala 2.10 finished with failure |
|
@jtgrabowski MaxConcurrentRequests is set only for the apply method of the query executor. The Write method is still throttled as before based on "max_concurrent_writes" I didn't want to add a new tuning parameter in a point release and the driver team told me this should be safe :) |
|
retest this please |
|
Build against Scala 2.10 finished with failure |
1 similar comment
|
Build against Scala 2.10 finished with failure |
|
Build against Scala 2.11 finished with failure |
|
Build against Scala 2.11 finished with success |
RussellSpitzer
force-pushed
the
SPARKC-503
branch
from
5ec5f7b
to
cd14ff2
Compare
|
Build against Scala 2.10 finished with success |
|
Build against Scala 2.11 finished with success |
| private val semaphore = new Semaphore(maxConcurrentTasks) | ||
| private val pendingFutures = new TrieMap[ListenableFuture[R], Boolean] | ||
| @volatile private var latestException: Option[Throwable] = None | ||
|
|
||
| def getLatestException(): Option[Throwable] = synchronized { |
There was a problem hiding this comment.
why do we synchronize here?
|
LGTM but please take a look at |
RussellSpitzer
force-pushed
the
SPARKC-503
branch
from
cd14ff2
to
7ec48b3
Compare
|
Build against Scala 2.10 finished with success |
|
Build against Scala 2.11 finished with success |
A change in the Cassandra Java driver has caused the executeAsync method
to become fully asynchronous. This means that executeAsync calls can
sometimes return an exception of BusyPoolException, meaning that queries
were issued and no open execution pathway could be found. These
exceptions are always retryable because they indicate the local driver
state is temporarly busy. So we will retry these exceptions.
To make sure we don't have too many retries we will also make sure to
limit the number of max concurrent requests to the current
"requestsPerConnection" as specified in the pooling options. We are
already setting the maxPool size to the number of cores per executor so
this should prohibit having more requests in flight than we have open
connections.