[
https://issues.apache.org/jira/browse/CASSANDRA-15399?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17062910#comment-17062910
]
David Capwell commented on CASSANDRA-15399:
-------------------------------------------
bq. "Reliable" message delivery alone in 15566 should make a large improvement..
How I wanted to do that was to dedupe requests on the recipients side, which
requires knowing what's going on (which is what this JIRA gave). You can do it
without this JIRA, but you would still need to do a lot of the heavy lifting of
this JIRA, which is why I liked them coupled.
bq. It requires all participants to be up and responsive through repair task
lifetime. It'd be nice to allow partial repair where some participants repair
successfully and it records failed(crashed/slow/message dropped) participants.
was planning to look into that after the transport and coordination was
reliable. There is some logic for if the the participants are down at the
start of the repair, but don't see anything about failing in the middle (other
than failing the repair); I know IR less than full but I assume both are the
same given the IR commit logic.
I see other open issues with repair that could make it more operational, stuff
like large ranges causing OOM, lack of cluster coordinator requiring external
coordinator (and everyone has their own), etc. My hope is to incrementally
improve this across several releases (I don't like giant patches).
> Add ability to track state in repair
> ------------------------------------
>
> Key: CASSANDRA-15399
> URL: https://issues.apache.org/jira/browse/CASSANDRA-15399
> Project: Cassandra
> Issue Type: Improvement
> Components: Consistency/Repair
> Reporter: David Capwell
> Assignee: David Capwell
> Priority: Normal
> Labels: pull-request-available
> Time Spent: 10m
> Remaining Estimate: 0h
>
> To enhance the visibility in repair, we should expose internal state via
> virtual tables; the state should include coordinator as well as participant
> state (validation, sync, etc.)
> I propose the following tables:
> repairs - high level summary of the global state of repair; this should be
> called on the coordinator.
> {code:sql}
> CREATE TABLE repairs (
> id uuid,
> keyspace_name text,
> table_names frozen<list<text>>,
> ranges frozen<list<text>>,
> coordinator text,
> participants frozen<list<text>>,
> state text,
> progress_percentage float,
> last_updated_at_millis bigint,
> duration_micro bigint,
> failure_cause text,
> PRIMARY KEY ( (id) )
> )
> {code}
> repair_tasks - represents RepairJob and participants state. This will show
> if validations are running on participants and the progress they are making;
> this should be called on the coordinator.
> {code:sql}
> CREATE TABLE repair_tasks (
> id uuid,
> session_id uuid,
> keyspace_name text,
> table_name text,
> ranges frozen<list<text>>,
> coordinator text,
> participant text,
> state text,
> state_description text,
> progress_percentage float, -- between 0.0 and 100.0
> last_updated_at_millis bigint,
> duration_micro bigint,
> failure_cause text,
> PRIMARY KEY ( (id), session_id, table_name, participant )
> )
> {code}
> repair_validations - shows the state of the validation task and updated
> periodically while validation is running; this should be called on the
> participants.
> {code:sql}
> CREATE TABLE repair_validations (
> id uuid,
> session_id uuid,
> ranges frozen<list<text>>,
> keyspace_name text,
> table_name text,
> initiator text,
> state text,
> progress_percentage float,
> queue_duration_ms bigint,
> runtime_duration_ms bigint,
> total_duration_ms bigint,
> estimated_partitions bigint,
> partitions_processed bigint,
> estimated_total_bytes bigint,
> failure_cause text,
> PRIMARY KEY ( (id), session_id, table_name )
> )
> {code}
> The main reason for exposing virtual tables rather than exposing through
> durable tables is to make sure what is exposed is accurate. In cases of
> write failures or node failures, the durable tables could become in-accurate
> and could add edge cases where the repair is not running but the tables say
> it is; by relying on repair's internal in-memory bookkeeping, these problems
> go away.
> This jira does not try to solve the following:
> 1) repair resiliency - there are edge cases where repair hits an error and
> runs forever (at least from nodetool's perspective).
> 2) repair stream tracking - I have not learned the streaming side yet and
> what I see is multiple implementations exist, so seems like high scope. My
> hope is to punt from this jira and tackle separately.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
