[
https://issues.apache.org/jira/browse/IGNITE-18203?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Alexander Lapin updated IGNITE-18203:
-------------------------------------
Summary: Table/index dependecy deadlock on node restart (was: Redesign
table/index creation to avoid dependency cycles)
> Table/index dependecy deadlock on node restart
> ----------------------------------------------
>
> Key: IGNITE-18203
> URL: https://issues.apache.org/jira/browse/IGNITE-18203
> Project: Ignite
> Issue Type: Improvement
> Reporter: Roman Puchkovskiy
> Assignee: Alexander Lapin
> Priority: Blocker
> Labels: ignite-3
> Fix For: 3.0.0-beta2
>
> Time Spent: 1h 40m
> Remaining Estimate: 0h
>
> Currently, index depends on a table, while table depends on RAFT groups
> start, which may need to apply commands (including inserts), and inserts
> depend on indices of the table. There is a dependency cycle. This is just one
> example of such a cycle in the current design.
> One idea is to split 'Table' entity into 2: 'Table Structure' and 'Full
> Table'. Table structure would only comprise objects representing bare table
> information (no indices), something like current
> {{{}TableImpl{}}}/{{{}InternalTable{}}}. Then, the following dependencies
> could be created:
> # Table Structure depends on Configuration
> # Index depends on Configuration and Table Structure
> # Sql Schema depends on Table Structure and Index
> # Full Table depends on Table Structure, Index, Sql Schema
> This requires a redesign of how {{{}TableManager{}}}, {{{}IndexManager{}}},
> {{SqlSchemaManager}} interact, probably {{TableManager}} needs to be broken
> in 2.
>
> *Upd*
> I've tried to redesign the TableManager in order to split the table into two
> parts: schema based "client" one (used by indeces, etc) and the one that
> assumes starting server endpoints. Unfortunately I didn't succeed in it. We
> will have one more attempt within distribution zone collocation related
> issues. Thus in order to fix the deadlock I've proposed rather crappy
> solution with async raft log application on restart based on the latch that
> bind PartitionReplicaListener.invoke and NodeImpl together. Ugly one that
> hopefully will be eliminated in the future.
> Ticket was renamed:
> "Redesign table/index creation to avoid dependency cycles" ->
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
