[ https://issues.apache.org/jira/browse/IGNITE-13052?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Ivan Rakov updated IGNITE-13052: -------------------------------- Description: Method reserveHistoryForExchange() is called on every partition map exchange. It's an expensive call: it requires iteration over the whole checkpoint history with possible retrieve of GroupState from WAL (it's stored on heap with SoftReference). On some deployments this operation can take several minutes. The idea of optimization is to calculate its result only on first PME (ideally, even before first PME, on recovery stage), keep resulting map (grpId, partId -> earlisetCheckpoint) on heap and update it if necessary. From the first glance, the map should be updated: 1) On checkpoint. If a new partition appears on local node, it should be registered in the map with current checkpoint. If a partition is evicted from local node, or changes its state to non-OWNING, it should be removed from the map. If checkpoint is marked as inapplicable for a certain group, the whole group should be removed from the map. 2) On checkpoint history cleanup. For every (grpId, partId), previous earliest checkpoint should be changed with setIfGreater to new earliest checkpoint. We should also extract WAL pointer reservation and filtering small partitions from reserveHistoryForExchange(), but this shouldn't be a problem. Another point: possibly, instead of storing map, we can keep earlistCheckpoint right in GridDhtLocalPartition. It may simplify implementation. Memory overhead of storing described map on heap in significant. Its size isn't greater than size of map returned from reserveHistoryForExchange(). Described fix should be much simpler than IGNITE-12429. was: Method reserveHistoryForExchange() is called on every partition map exchange. It's an expensive call: it requires iteration over the whole checkpoint history with possible retrieve of GroupState from WAL (it's stored on heap with SoftReference). On some deployments this operation can take several minutes. The idea of optimization is to calculate its result only on first PME (ideally, even before first PME, on recovery stage), keep resulting map (grpId, partId -> earlisetCheckpoint) on heap and update it if necessary. From the first glance, the map should be updated: 1) On checkpoint. If a new partition appears on local node, it should be registered in the map with current checkpoint. If a partition is evicted from local node, or changes its state to non-OWNING, it should be removed from the map. If checkpoint is marked as inapplicable for a certain group, the whole group should be removed from the map. 2) On checkpoint history cleanup. For every (grpId, partId), previous earliest checkpoint should be changed with setIfGreater to new earliest checkpoint. We should also extract WAL pointer reservation and filtering small partitions from reserveHistoryForExchange(), but this shouldn't be a problem. Another point: possibly, instead of storing map, we can keep earlistCheckpoint right in GridDhtLocalPartition. It may simplify implementation. Memory overhead of storing described map on heap in significant. It's size isn't greater than size of map returned from reserveHistoryForExchange(). Described fix should be much simpler than IGNITE-12429. > Calculate result of reserveHistoryForExchange in advance > -------------------------------------------------------- > > Key: IGNITE-13052 > URL: https://issues.apache.org/jira/browse/IGNITE-13052 > Project: Ignite > Issue Type: Improvement > Reporter: Ivan Rakov > Priority: Major > Original Estimate: 80h > Remaining Estimate: 80h > > Method reserveHistoryForExchange() is called on every partition map exchange. > It's an expensive call: it requires iteration over the whole checkpoint > history with possible retrieve of GroupState from WAL (it's stored on heap > with SoftReference). On some deployments this operation can take several > minutes. > The idea of optimization is to calculate its result only on first PME > (ideally, even before first PME, on recovery stage), keep resulting map > (grpId, partId -> earlisetCheckpoint) on heap and update it if necessary. > From the first glance, the map should be updated: > 1) On checkpoint. If a new partition appears on local node, it should be > registered in the map with current checkpoint. If a partition is evicted from > local node, or changes its state to non-OWNING, it should be removed from the > map. If checkpoint is marked as inapplicable for a certain group, the whole > group should be removed from the map. > 2) On checkpoint history cleanup. For every (grpId, partId), previous > earliest checkpoint should be changed with setIfGreater to new earliest > checkpoint. > We should also extract WAL pointer reservation and filtering small partitions > from reserveHistoryForExchange(), but this shouldn't be a problem. > Another point: possibly, instead of storing map, we can keep > earlistCheckpoint right in GridDhtLocalPartition. It may simplify > implementation. > Memory overhead of storing described map on heap in significant. Its size > isn't greater than size of map returned from reserveHistoryForExchange(). > Described fix should be much simpler than IGNITE-12429. -- This message was sent by Atlassian Jira (v8.3.4#803005)