/couchbase-performance-tuning

name: couchbase-performance-tuning
description: "Diagnose and tune cluster-level performance problems in Couchbase. Use whenever the user asks about slow KV operations, high latency, low throughput, DCP backpressure, connection limits, thread pool tuning, compaction impact on performance, autocompaction settings, KV engine tuning, vBucket distribution, rebalance performance, disk I/O bottlenecks, CPU saturation on Couchbase nodes, network throughput limits, or 'my cluster is healthy but slow.' Distinct from couchbase-sqlpp-tuning (SQL++ query tuning — index design and EXPLAIN plans), couchbase-observability (what metrics to watch), and couchbase-sizing (how much capacity to provision). This skill covers the operational tuning layer: what to change when the cluster is right-sized but still not performing."
license: MIT

Couchbase Performance Tuning

A skill for diagnosing and fixing performance problems at the cluster level — KV latency, throughput limits, disk I/O, compaction, connection saturation, and thread pool configuration.

Distinct from:

• `couchbase-sqlpp-tuning` — SQL++ query tuning (index design, EXPLAIN plans, anti-patterns)
• `couchbase-observability` — what metrics to monitor and alert thresholds
• `couchbase-sizing` — how much capacity to provision in the first place

If the question is "my queries are slow," go to `couchbase-sqlpp-tuning`. If the question is "my cluster is right-sized but everything is slow," this is the right skill.

When this skill applies

• "KV get/set latency is higher than expected"
• "Throughput isn't reaching the hardware's capability"
• "Compaction is killing performance"
• "We're hitting connection limits"
• "Rebalance is taking too long"
• "High CPU on Couchbase nodes but no obvious cause"
• "Disk I/O is spiking unpredictably"
• "DCP consumers are falling behind"

Pick the right reference

| Question | Read | |---|---| | "KV latency / throughput — diagnosis and tuning" | `references/kv-tuning.md` | | "Compaction — autocompaction settings, impact, tuning" | `references/compaction.md` | | "Connection limits, thread pools, OS-level tuning" | `references/system-tuning.md` |

The diagnosis sequence

Before tuning anything, locate the actual bottleneck:

1. **Is it memory?** Check `ep_mem_used / ep_mem_high_wat`. If > 85%, ejections are happening and reads go to disk. Fix: add RAM, add nodes, or reduce working set.

2. **Is it disk I/O?** Check `ep_bg_fetched` (reads going to disk) and `ep_diskqueue_drain` vs `ep_diskqueue_fill`. Fix: faster storage, Magma (if on 8.0), or reduce write rate.

3. **Is it CPU?** Check per-node CPU utilization. Which service is consuming it? Query and Index are CPU-heavy; KV should be low-CPU unless you're near capacity. Fix: dedicated nodes per service, or add nodes.

4. **Is it network?** Check `bytes_sent` and `bytes_received` per node against the node's NIC capacity. Fix: higher-bandwidth instances, or reduce replication/XDCR traffic.

5. **Is it compaction?** Check if high disk I/O correlates with compaction windows. Fix: adjust compaction schedule, thresholds, or parallelism.

6. **Is it connection count?** Check `curr_connections` per node. Fix: connection pooling in SDK, reduce max connections per pool.

Only tune after locating the bottleneck. Tuning the wrong thing wastes time and can make things worse.

Related skills

• `couchbase-sqlpp-tuning` — query-level performance (EXPLAIN, indexes, CBO)
• `couchbase-observability` — key metrics definitions and alert thresholds
• `couchbase-magma` — Magma storage engine characteristics that affect disk I/O patterns
• `couchbase-sizing` — if tuning can't solve the problem, the next step is adding capacity