[
https://issues.apache.org/jira/browse/IGNITE-21115?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Roman Puchkovskiy updated IGNITE-21115:
---------------------------------------
Description:
When an index is created (and appears in the REGISTERED state), we must wait
for all transactions started on schemas before the index has appeared to finish.
Primary replica of the partition zero of the table owning the created index is
responsible for the wait.
# When a node sees appearance of a REGISTERED index state in the Catalog, and
it’s a primary of partition zero of the table owning the index, it starts the
waiter
# When a node becomes the primary of the partition zero of a table, and the
table contains indices that are in the REGISTERED state (at the moment of the
primary lease start), it starts a waiter per each such index
# When a node ceases to be the primary of the partition zero of a table, it
stops all the waiters it started
# When the waiter finishes its waiting and the node is still a primary, it
switches the index to the BACKFILLING state
h5. Waiter operation
# Wait for HLC.now() to become greater than ActivationMoment(index REGISTERED
state) + MaxClockSkew (to make sure that the REGISTERED state activates on each
node in the cluster)
# Take all nodes in the cluster Logical Topology (LT) from the CMG Leader (not
from the local state delivered by CMG learners) and put it in the *Nodes* list.
({*}Nodes{*} contains all the nodes that might have started RW transactions on
schemas before the index appeared; other nodes had either left the topology, or
they can only coordinate RW transactions started on schema versions containing
the index)
Poll each node in the *Nodes* list until either its RwTransactionsFinished(V)
(see IGNITE-21112) returns true (where V is the catalog version where the index
had appeared) or the node leaves the LT (if the node leaves the Physical
Topology, we proceed waiting till it falls off the LT).
After the waiter finishes, we must switch the index to the BACKFILLING state.
was:
When an index is created (and appears in the REGISTERED state), we must wait
for all transactions started on schemas before the index has appeared to finish.
Primary replica of the partition zero of the table owning the created index is
responsible for the wait.
# When a node sees appearance of a REGISTERED index state in the Catalog, and
it’s a primary of partition zero of the table owning the index, it starts the
waiter
# When a node becomes the primary of the partition zero of a table, and the
table contains indices that are in the REGISTERED state (at the moment of the
primary lease start), it starts a waiter per each such index
# When a node ceases to be the primary of the partition zero of a table, it
stops all the waiters it started
# When the waiter finishes its waiting and the node is still a primary, it
switches the index to the BACKFILLING state
h5. Waiter operation
# Wait for HLC.now() to become greater than ActivationMoment(index REGISTERED
state) + MaxClockSkew (to make sure that the REGISTERED state activates on each
node in the cluster)
# Take all nodes in the cluster Logical Topology (LT) from the CMG Leader (not
from the local state delivered by CMG learners) and put it in the *Nodes* list.
({*}Nodes{*} contains all the nodes that might have started RW transactions on
schemas before the index appeared; other nodes had either left the topology, or
they can only coordinate RW transactions started on schema versions containing
the index)
Poll each node in the *Nodes* list until either its RwTransactionsFinished(V)
returns true (where V is the catalog version where the index had appeared) or
the node leaves the LT (if the node leaves the Physical Topology, we proceed
waiting till it falls off the LT).
After the waiter finishes, we must switch the index to the BACKFILLING state.
> Wait for completion of transactions started before index appearance
> -------------------------------------------------------------------
>
> Key: IGNITE-21115
> URL: https://issues.apache.org/jira/browse/IGNITE-21115
> Project: Ignite
> Issue Type: Improvement
> Reporter: Roman Puchkovskiy
> Priority: Major
> Labels: ignite-3
> Fix For: 3.0.0-beta2
>
>
> When an index is created (and appears in the REGISTERED state), we must wait
> for all transactions started on schemas before the index has appeared to
> finish.
> Primary replica of the partition zero of the table owning the created index
> is responsible for the wait.
> # When a node sees appearance of a REGISTERED index state in the Catalog,
> and it’s a primary of partition zero of the table owning the index, it starts
> the waiter
> # When a node becomes the primary of the partition zero of a table, and the
> table contains indices that are in the REGISTERED state (at the moment of the
> primary lease start), it starts a waiter per each such index
> # When a node ceases to be the primary of the partition zero of a table, it
> stops all the waiters it started
> # When the waiter finishes its waiting and the node is still a primary, it
> switches the index to the BACKFILLING state
> h5. Waiter operation
> # Wait for HLC.now() to become greater than ActivationMoment(index
> REGISTERED state) + MaxClockSkew (to make sure that the REGISTERED state
> activates on each node in the cluster)
> # Take all nodes in the cluster Logical Topology (LT) from the CMG Leader
> (not from the local state delivered by CMG learners) and put it in the
> *Nodes* list. ({*}Nodes{*} contains all the nodes that might have started RW
> transactions on schemas before the index appeared; other nodes had either
> left the topology, or they can only coordinate RW transactions started on
> schema versions containing the index)
> Poll each node in the *Nodes* list until either its RwTransactionsFinished(V)
> (see IGNITE-21112) returns true (where V is the catalog version where the
> index had appeared) or the node leaves the LT (if the node leaves the
> Physical Topology, we proceed waiting till it falls off the LT).
> After the waiter finishes, we must switch the index to the BACKFILLING state.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
