[
https://issues.apache.org/jira/browse/KUDU-3801?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Gabriella Lotz updated KUDU-3801:
---------------------------------
Description:
The auto-rebalancer and auto-leader-rebalancer have been in the tree behind
experimental flags for a while, but there's no user-facing documentation.
`docs/administration.adoc` only covers the CLI `kudu cluster rebalance` tool,
and `docs/background_tasks.adoc` doesn't mention either task. The only
references I could find in `docs/` are stale entries under
`prior_release_notes.adoc`.
Add a new section, I think under `docs/background_tasks.adoc` since these are
master background tasks, covering:
- What each task does and how they relate to each other, and to the CLI `kudu
cluster rebalance` tool already documented in `administration.adoc`.
- Prerequisites: `--raft_prepare_replacement_before_eviction=true` (3-4-3) is
required for `auto_rebalancing_enabled` and enforced by the group validator at
`catalog_manager.cc:570`. Whether it's needed for
`auto_leader_rebalancing_enabled` is being decided under the related sub-task;
reflect the outcome here.
- Operator-visible flags with recommended values and when to change them:
`auto_rebalancing_interval_seconds`, `auto_rebalancing_max_moves_per_server`,
`auto_rebalancing_load_imbalance_threshold`,
`auto_rebalancing_prefer_follower_replica_moves`,
`auto_rebalancing_enable_range_rebalancing`,
`auto_leader_rebalancing_interval_seconds`,
`leader_rebalancing_max_moves_per_round`,
`leader_rebalancing_ignore_soft_deleted_tables`.
- The metrics that were added: `auto_rebalancer_leader_moves_scheduled`,
`auto_rebalancer_follower_moves_scheduled`, `auto_rebalancer_rounds_completed`,
the four counters added under KUDU-3791, and the leader-skew gauge added under
the related sub-task.
- Behavior with tservers in maintenance mode: destinations in
`MAINTENANCE_MODE` are skipped, and sources shed replicas/leaders.
- Interaction with the CLI rebalancer: they share the same replica-move
machinery, so running both at once isn't recommended.
Also add a release-notes entry in `docs/release_notes.adoc` announcing that the
two flags are no longer experimental, with a pointer to the new section
> Add documentation
> -----------------
>
> Key: KUDU-3801
> URL: https://issues.apache.org/jira/browse/KUDU-3801
> Project: Kudu
> Issue Type: Sub-task
> Reporter: Gabriella Lotz
> Priority: Major
>
> The auto-rebalancer and auto-leader-rebalancer have been in the tree behind
> experimental flags for a while, but there's no user-facing documentation.
> `docs/administration.adoc` only covers the CLI `kudu cluster rebalance` tool,
> and `docs/background_tasks.adoc` doesn't mention either task. The only
> references I could find in `docs/` are stale entries under
> `prior_release_notes.adoc`.
> Add a new section, I think under `docs/background_tasks.adoc` since these are
> master background tasks, covering:
> - What each task does and how they relate to each other, and to the CLI `kudu
> cluster rebalance` tool already documented in `administration.adoc`.
> - Prerequisites: `--raft_prepare_replacement_before_eviction=true` (3-4-3) is
> required for `auto_rebalancing_enabled` and enforced by the group validator
> at `catalog_manager.cc:570`. Whether it's needed for
> `auto_leader_rebalancing_enabled` is being decided under the related
> sub-task; reflect the outcome here.
> - Operator-visible flags with recommended values and when to change them:
> `auto_rebalancing_interval_seconds`, `auto_rebalancing_max_moves_per_server`,
> `auto_rebalancing_load_imbalance_threshold`,
> `auto_rebalancing_prefer_follower_replica_moves`,
> `auto_rebalancing_enable_range_rebalancing`,
> `auto_leader_rebalancing_interval_seconds`,
> `leader_rebalancing_max_moves_per_round`,
> `leader_rebalancing_ignore_soft_deleted_tables`.
> - The metrics that were added: `auto_rebalancer_leader_moves_scheduled`,
> `auto_rebalancer_follower_moves_scheduled`,
> `auto_rebalancer_rounds_completed`, the four counters added under KUDU-3791,
> and the leader-skew gauge added under the related sub-task.
> - Behavior with tservers in maintenance mode: destinations in
> `MAINTENANCE_MODE` are skipped, and sources shed replicas/leaders.
> - Interaction with the CLI rebalancer: they share the same replica-move
> machinery, so running both at once isn't recommended.
> Also add a release-notes entry in `docs/release_notes.adoc` announcing that
> the two flags are no longer experimental, with a pointer to the new section
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
