[
https://issues.apache.org/jira/browse/IGNITE-23159?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Denis Chudov updated IGNITE-23159:
----------------------------------
Description:
Let's assume a table:
{noformat}
Goods:
id int primary key
name: varchar
category: varchar, indexed with non-unique index
{noformat}
and the following scenario, where two transaction are executed concurrently on
nodes A and B, and primary keys 1 and 2 will be located on these nodes
respectively:
A: tx1: insert (1, Apple, Fruits); // takes IX lock
B: tx2: scan(Fruits) // found nothing, locks nothing
B: tx1: insert(2, Peach, Fruits); // takes IX lock
A: tx2: scan(Fruits) // waits for IX lock release to take S lock
tx1:commit
A: tx2: unlocked and reads Apple by index and finished
Serializability is broken, because tx2 didn't see Peach and was ordered later,
but did see Apple.
This needs to be fixed by taking index locks both on nodes A and B (actually on
all primary replicas), with waiting for all nodes to ack to lock acquisition on
the index before inserting the value.
This will have big performance impact, until is addressed by implementing
separate distribution zones for indexes (requires a design).
The same issue is applied to sorted indexes.
was:
Let's assume a table:
{noformat}
Goods:
id int primary key
name: varchar
category: varchar, indexed with non-unique index
{noformat}
and the following scenario, where two transaction are executed concurrently on
nodes A and B, and primary keys 1 and 2 will be located on these nodes
respectively:
A: tx1: insert (1, Apple, Fruits); // takes IX lock
B: tx2: scan(Fruits) // found nothing, locks nothing
B: tx1: insert(2, Peach, Fruits); // takes IX lock
A: tx2: scan(Fruits) // waits for IX lock release to take S lock
tx1:commit
A: tx2: unlocked and reads Apple by index and finished
Serializability is broken, because tx2 didn't see Peach and was ordered later,
but did see Apple.
This needs to be fixed by taking index locks both on nodes A and B (actually on
all primary replicas).
This will have big performance impact, until is addressed by implementing
separate distribution zones for indexes (requires a design).
The same issue is applied to sorted indexes.
> Fix broken serializability of transactions on tables having secondary indexes
> -----------------------------------------------------------------------------
>
> Key: IGNITE-23159
> URL: https://issues.apache.org/jira/browse/IGNITE-23159
> Project: Ignite
> Issue Type: Bug
> Reporter: Alexey Scherbakov
> Priority: Major
> Labels: ignite-3
> Fix For: 3.0
>
>
> Let's assume a table:
> {noformat}
> Goods:
> id int primary key
> name: varchar
> category: varchar, indexed with non-unique index
> {noformat}
> and the following scenario, where two transaction are executed concurrently
> on nodes A and B, and primary keys 1 and 2 will be located on these nodes
> respectively:
> A: tx1: insert (1, Apple, Fruits); // takes IX lock
> B: tx2: scan(Fruits) // found nothing, locks nothing
> B: tx1: insert(2, Peach, Fruits); // takes IX lock
> A: tx2: scan(Fruits) // waits for IX lock release to take S lock
> tx1:commit
> A: tx2: unlocked and reads Apple by index and finished
> Serializability is broken, because tx2 didn't see Peach and was ordered
> later, but did see Apple.
> This needs to be fixed by taking index locks both on nodes A and B (actually
> on all primary replicas), with waiting for all nodes to ack to lock
> acquisition on the index before inserting the value.
> This will have big performance impact, until is addressed by implementing
> separate distribution zones for indexes (requires a design).
> The same issue is applied to sorted indexes.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
