[
https://issues.apache.org/jira/browse/CASSANDRA-17401?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17809755#comment-17809755
]
Long Pan commented on CASSANDRA-17401:
--------------------------------------
[~chovatia.jayd...@gmail.com] and I managed to reproduce the issue. Basically a
large number of QPS and client connections are necessary to reproduce it. Here
are the steps to reproduce:
*Server Setup:*
A 3-node Cassandra cluster. Each node with 64GB mem (16GB heap), 7 CPU cores.
({{{}native_transport_max_threads = 1024){}}}
*Keyspace/Table:*
CREATE KEYSPACE test_ks WITH REPLICATION = \{ ‘class’ :
‘NetworkTopologyStrategy’, ‘datacenter1’ : 3 } ;
CREATE TABLE test_ks.table1 ( p_id text, c_id text, v text);
*Client Setup:*
30 hosts. Each host run the following pseudo-code, using *GoCql* client:
cluster.CQLVersion = "3.4.0"
cluster.ProtoVersion = 4
cluster.Timeout = 5s
cluster.ConnectTimeout = 10s
cluster.NumConns = 3
cluster.Consistency = LocalQuorum
cluster.RetryPolicy = SimpleRetryPolicy\{NumRetries: 1}
cluster.SocketKeepalive = 20s
cluster.HostSelectionPolicy = RoundRobinHostPolicy
sessionCount = 30
qpsPerSession = 30
cqlQuery = "SELECT p_id,c_id,v FROM test_ks.table1 WHERE p_id = ? AND c_id
= ?"
for (i = 0; i < sessionCount; i++) \{
session = cluster.createSession
rateLimiter = NewRateLimiter(qpsPerSession)
newGoRoutine.run( sendReads(session, rateLimiter) )
}
/ *
sendReads(session, rateLimiter) \{
for {
newGoRoutine.run (
if (rateLimiter.allow) {
session.execute(cqlQuery, randomString, randomString)
}
)
}
}
*/
Traffic generated this way will result in ~10K coordiator QPS and ~3k client
connections per Cassandra node.
*Trigger Point:*
Manually issue a CQL query to add a column in the table: “ALTER TABLE
test_ks.table1 ADD new_col text;”
{*}Symmpton{*}:
Seconds after the trigger point, one or more Cassandra nodes will show number
of native_transport threads reaching {{{}native_transport_max_threads{}}}, and
pending native transport tasks grow endlessly.
> Race condition in QueryProcessor causes just prepared statement not to be in
> the prepared statements cache
> ----------------------------------------------------------------------------------------------------------
>
> Key: CASSANDRA-17401
> URL: https://issues.apache.org/jira/browse/CASSANDRA-17401
> Project: Cassandra
> Issue Type: Bug
> Reporter: Ivan Senic
> Priority: Normal
> Time Spent: 10m
> Remaining Estimate: 0h
>
> The changes in the
> [QueryProcessor#prepare|https://github.com/apache/cassandra/blame/cassandra-4.0.2/src/java/org/apache/cassandra/cql3/QueryProcessor.java#L575-L638]
> method that were introduced in versions *4.0.2* and *3.11.12* can cause a
> race condition between two threads trying to concurrently prepare the same
> statement. This race condition can cause removing of a prepared statement
> from the cache, after one of the threads has received the result of the
> prepare and eventually uses MD5Digest to call
> [QueryProcessor#getPrepared|https://github.com/apache/cassandra/blame/cassandra-4.0.2/src/java/org/apache/cassandra/cql3/QueryProcessor.java#L212-L215].
> The race condition looks like this:
> * Thread1 enters _prepare_ method and resolves _safeToReturnCached_ as false
> * Thread1 executes eviction of hashes
> * Thread2 enters _prepare_ method and resolves _safeToReturnCached_ as false
> * Thread1 prepares the statement and caches it
> * Thread1 returns the result of the prepare
> * Thread2 executes eviction of hashes
> * Thread1 tries to execute the prepared statement with the received
> MD5Digest, but statement is not in the cache as it was evicted by Thread2
> I tried to reproduce this by using a Java driver, but hitting this case from
> a client side is highly unlikely and I can not simulate the needed race
> condition. However, we can easily reproduce this in Stargate (details
> [here|https://github.com/stargate/stargate/pull/1647]), as it's closer to
> QueryProcessor.
> Reproducing this in a unit test is fairly easy. I am happy to showcase this
> if needed.
> Note that the issue can occur only when safeToReturnCached is resolved as
> false.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
