[
https://issues.apache.org/jira/browse/IGNITE-22286?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17878622#comment-17878622
]
Denis Chudov edited comment on IGNITE-22286 at 9/6/24 10:31 AM:
----------------------------------------------------------------
I did a benchmark: explicit tx with one upsert, high contention on keys. Tx is
rolled back if upsert fails. Upsert duration measured from upsert() call to
returning control to user. Time is in millis.
Code:
https://github.com/gridgain/apache-ignite-3/commit/b089f94fc4470efdead10f6e77a9b9a105a5a3f2
Benchmark results:
{code:java}
+------------------------------------+-------------+------------------+-------------------------+
| Configuration | total txns | rolled back txns | average
upsert duration |
+------------------------------------+-------------+------------------+-------------------------+
| Timeout wait, 1 node, 2 cores| 1611 | 166 |
7.583306 |
| Timeout wait, 3 nodes, 2 cores| 1705 | 120 |
17.575059 |
| Timeout wait, 1 node, 8 cores| 3445 | 394 |
3.911635 |
| Timeout wait, 3 nodes, 8 cores| 3055 | 371 |
4.813581 |
| WaitDie + retries, 1 node, 2 cores| 8160 | 169 |
12.158225 |
| WaitDie + retries, 3 nodes, 2 cores| 3853 | 213 |
10.584286 |
| WaitDie + retries, 1 node, 8 cores| 22797 | 2477 |
8.767905 |
| WaitDie + retries, 3 nodes, 8 cores| 13092 | 1285 |
12.519806 |
+------------------------------------+-------------+------------------+-------------------------+
{code}
was (Author: denis chudov):
I did a benchmark: explicit tx with one upsert, high contention on keys. Tx is
rolled back if upsert fails. Upsert duration measured from upsert() call to
returning control to user. Time is in millis.
Benchmark results:
{code:java}
+------------------------------------+-------------+------------------+-------------------------+
| Configuration | total txns | rolled back txns | average
upsert duration |
+------------------------------------+-------------+------------------+-------------------------+
| Timeout wait, 1 node, 2 cores| 1611 | 166 |
7.583306 |
| Timeout wait, 3 nodes, 2 cores| 1705 | 120 |
17.575059 |
| Timeout wait, 1 node, 8 cores| 3445 | 394 |
3.911635 |
| Timeout wait, 3 nodes, 8 cores| 3055 | 371 |
4.813581 |
| WaitDie + retries, 1 node, 2 cores| 8160 | 169 |
12.158225 |
| WaitDie + retries, 3 nodes, 2 cores| 3853 | 213 |
10.584286 |
| WaitDie + retries, 1 node, 8 cores| 22797 | 2477 |
8.767905 |
| WaitDie + retries, 3 nodes, 8 cores| 13092 | 1285 |
12.519806 |
+------------------------------------+-------------+------------------+-------------------------+
{code}
> Remove waitTimeout in deadlock prevention policy
> ------------------------------------------------
>
> Key: IGNITE-22286
> URL: https://issues.apache.org/jira/browse/IGNITE-22286
> Project: Ignite
> Issue Type: Improvement
> Affects Versions: 3.0.0-beta1
> Reporter: Alexey Scherbakov
> Assignee: Denis Chudov
> Priority: Major
> Labels: ignite-3
>
> After IGNITE-21540 and IGNITE-20127 we now have proper retries on client side.
> This means we no longer need
> org.apache.ignite.internal.tx.DeadlockPreventionPolicy#waitTimeout as a part
> of deadock prevention policy.
> Moreover, client side retries has benefit in the following scenario (having
> in mind WAIT_DIE prevention):
> # tx1 takes lock at timestamp 10
> # tx2 tries to take lock at timestamp 20 and goes for retry (without holding
> lock)
> # tx1 lock is released
> # tx3 takes lock at timestamp 30
> # tx3 lock is released
> # tx2 attemps to lock after retry and succeeds
> Without retry (without holding locks) on step 2 tx3 would retry too on step 4.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
