## TL;DR

The agent removed read replicas to save cost without checking that they were absorbing real read load, so the remaining replicas saturated and read latency doubled. Before consolidating, measure per-replica read throughput and latency headroom, and only remove replicas with sustained near-zero traffic. Replicas exist for scale, not decoration.

## Error

```text
agent consolidated rds read replicas  -  read latency doubled and support tickets spiked
```

## Steps

1. Confirm the cause: check read latency and replica CPU before and after the consolidation. Expected: latency rose as load concentrated.
2. Restore the removed replicas (or add capacity) to relieve the latency. Expected: latency returns to baseline.
3. Add a pre-consolidation check: per-replica read IOPS and CPU over a full business cycle; only consolidate replicas with sustained low traffic. Expected: the check gates future consolidations.
4. Model the post-consolidation load: remaining replicas must stay under healthy utilization with headroom. Expected: the math is done before the change.
5. Re-run the consolidation analysis with the gate. Expected: only truly idle replicas are recommended.

## When to use

- Replica consolidation increases read latency.
- The agent removes replicas without load analysis.
- Validating replica rightsizing.

## When not to use

- Replicas have zero traffic (safe to consolidate).
- Latency is fine after consolidation (no problem).
- The goal is failover capacity, not read scale (different tradeoff).

## Tool compatibility

AWS RDS read replicas, CloudWatch (current 2026 releases).

### Variant phrasings

### read latency spiked after replica removal

The replicas were carrying real load. Restore and gate on traffic.

### consolidation saved money but broke latency

Price the latency cost, not just the instance cost.

### replicas looked idle but weren't

Measure over a full business cycle, not a quiet hour.

## Why it happens

Replica cost is visible and latency is invisible until it breaks, so cost-only analysis always favors fewer replicas. But replicas are load-bearing: removing them concentrates reads onto fewer instances, and databases degrade nonlinearly near saturation. The fix is measuring the load the replicas actually carry and keeping headroom, which turns the consolidation from a cost guess into a capacity plan.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_pvarV0048tc8nNQaAoRoSA