การจัดสรรชาร์ดและการปรับสมดุลใหม่

When you create an index or add a node, Elasticsearch must decide which node holds each shard. This process is shard allocation, and keeping it balanced is key to a healthy, performant cluster.

The master node runs allocation deciders for every unassigned shard. They check disk space, shard count, awareness rules, and filters before placing a shard. Failing any decider keeps the shard UNASSIGNED.

A replica is never allocated to the same node as its primary. This guarantees that losing one node cannot take out both copies of the same data. With one replica you survive any single-node failure.

Use the _cat/shards API to see where each shard sits and its state. STARTED means active; RELOCATING means it is moving.

When a shard will not allocate, the _cluster/allocation/explain API tells you exactly which decider blocked it and why. This is the first stop for debugging a yellow or red cluster.

Allocation awareness spreads copies across physical boundaries like racks or availability zones. You tag nodes with custom attributes and tell the cluster to be aware of them.

Forced awareness goes further: it refuses to over-allocate replicas to surviving zones if a zone goes down, preventing one zone from holding all copies. This protects capacity planning during outages.

You can pin or exclude an index from specific nodes using allocation filters. This is handy for hot/warm architectures where older indices migrate to cheaper hardware.

After nodes join or leave, the cluster rebalances to even out shard counts. This is automatic but generates network and disk I/O, so it should be controlled during sensitive operations.

Before a rolling restart, disable allocation so the cluster does not waste effort relocating shards from a node you are about to bring back. Re-enable it afterward.

Limit concurrent recoveries with cluster.routing.allocation.node_concurrent_recoveries and bandwidth with indices.recovery.max_bytes_per_sec to keep rebalancing from starving live queries.

You learned how shards are placed and balanced:

• Allocation deciders choose shard locations; replicas avoid their primary's node.
• Debug stuck shards with allocation/explain.
• Allocation awareness and forced awareness spread copies across zones.
• Use allocation filters for hot/warm tiers and disable allocation during rolling restarts.