[
https://issues.apache.org/jira/browse/IGNITE-13358?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Alexey Scherbakov updated IGNITE-13358:
---------------------------------------
Description:
We have several issues related to a partition clearing worth fixing.
1. PartitionsEvictManager doent's provide obvious guarantees for a correctness
when a node or a cache group is stopped while partitions are concurrently
clearing.
2. GridDhtLocalPartition#awaitDestroy is called while holding topology write
lock, which is deadlock prone, because we currently require write lock to
destroy a partition.
3. GridDhtLocalPartition contains a lot of messy code related to partition
clearing, most notably ClearFuture, but the clearing is done by
PartitionsEvictManager. We should get rid of a clearing code in
GridDhtLocalPartition. This should also bring better code readility and help
understand what happening during a clearing.
4. Currently moving partitions are cleared before rebalancing in the order
different to rebalanceOrder, breaking the contract.
5. The clearing logic for for moving partitions (before rebalancing) seems
incorrect: it's possible to lost updates received during clearing.
6. To clear partitions before full rebalancing we utilize same threads as for a
partition eviction. This can slow rebalancing even if we have resources. Better
to clear partitions in the rebalance pool (explicitely dedicated by user).
7. It's possible to reserve a renting partition, which have absolutely no
meaning. All operations with a renting partitions (except clearing) are a waste
of resources.
8. Partition eviction causes system pool starvation if a number of thread in
system pool is < 8. This can break crucial functionality.
was:
We have several issues related to a partition clearing worth fixing.
1. PartitionsEvictManager doent's provide obvious guarantees for a correctness
when a node or a cache group is stopped while partitions are concurrently
clearing.
2. GridDhtLocalPartition#awaitDestroy is called while holding topology write
lock, which is deadlock prone, because we currently require write lock to
destroy a partition.
3. GridDhtLocalPartition contains a lot of messy code related to partition
clearing, most notably ClearFuture, but the clearing is done by
PartitionsEvictManager. We should get rid of a clearing code in
GridDhtLocalPartition. This should also bring better code readility and help
understand what happening during a clearing.
4. Currently moving partitions are cleared before rebalancing in the order
different to rebalanceOrder, breaking the contract.
5. The clearing logic for for moving partitions (before rebalancing) seems
incorrect: it's possible to lost updates received during clearing.
6. To clear partitions before full rebalancing we utilize same threads as for a
partition eviction. This can slow rebalancing even if we have resources. Better
to clear partitions in the rebalance pool (explicitely dedicated by user).
7. It's possible to reserve a renting partition, which have absolutely no
meaning. All operations with a renting partitions (except clearing) are a waste
of resources.
> Improvements for partition clearing related parts
> -------------------------------------------------
>
> Key: IGNITE-13358
> URL: https://issues.apache.org/jira/browse/IGNITE-13358
> Project: Ignite
> Issue Type: Improvement
> Reporter: Alexey Scherbakov
> Assignee: Alexey Scherbakov
> Priority: Major
>
> We have several issues related to a partition clearing worth fixing.
> 1. PartitionsEvictManager doent's provide obvious guarantees for a
> correctness when a node or a cache group is stopped while partitions are
> concurrently clearing.
> 2. GridDhtLocalPartition#awaitDestroy is called while holding topology write
> lock, which is deadlock prone, because we currently require write lock to
> destroy a partition.
> 3. GridDhtLocalPartition contains a lot of messy code related to partition
> clearing, most notably ClearFuture, but the clearing is done by
> PartitionsEvictManager. We should get rid of a clearing code in
> GridDhtLocalPartition. This should also bring better code readility and help
> understand what happening during a clearing.
> 4. Currently moving partitions are cleared before rebalancing in the order
> different to rebalanceOrder, breaking the contract.
> 5. The clearing logic for for moving partitions (before rebalancing) seems
> incorrect: it's possible to lost updates received during clearing.
> 6. To clear partitions before full rebalancing we utilize same threads as for
> a partition eviction. This can slow rebalancing even if we have resources.
> Better to clear partitions in the rebalance pool (explicitely dedicated by
> user).
> 7. It's possible to reserve a renting partition, which have absolutely no
> meaning. All operations with a renting partitions (except clearing) are a
> waste of resources.
> 8. Partition eviction causes system pool starvation if a number of thread in
> system pool is < 8. This can break crucial functionality.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
