## TL;DR
GetMetricData allows 500 metrics per call; rate exceeded means you are calling too often with too few metrics per call. Batch aggressively: pack hundreds of metric queries into each call, cache results for the metric period, and stop polling faster than the metric resolution. Most rate problems come from one-metric-per-call patterns that a single batched call replaces.

## The query
```text
CloudWatch "Rate exceeded" on GetMetricData: how to batch requests
```

## Use this when
- GetMetricData returns rate exceeded errors
- Metric collection is chatty
- Designing CloudWatch-based monitoring
- After adding many new metrics

## Not for when
- Other CloudWatch API rate limits (logs, etc.)
- Metric math expression errors
- Missing metrics (different issue)

## Steps

### Step 1: Find the chatty callers
Identify which code calls GetMetricData and with how many metrics per call. One metric per call in a loop is the classic pattern; it burns the quota hundreds of times faster than batched calls.
Expected output: the callers and their batch sizes mapped.

### Step 2: Batch to hundreds of metrics per call
Restructure calls to pack up to 500 metric queries per GetMetricData request. This single change usually eliminates rate errors entirely.
Expected output: call count dropping by orders of magnitude.

### Step 3: Cache at the metric period
Cache results for at least the metric's period (usually 60s or 300s). Polling faster than the data refreshes wastes quota for identical numbers.
Expected output: no duplicate queries within a period.

### Step 4: Add backoff with jitter anyway
Even batched callers hit limits during bursts. Implement exponential backoff with jitter so the calls you do make degrade gracefully under throttling.
Expected output: throttling absorbed without errors surfacing.

### Step 5: Consider metric streams for high volume
For very high metric volumes, CloudWatch metric streams (Kinesis Firehose delivery) replace polling entirely. Polling does not scale to thousands of metrics; streams do.
Expected output: the collection architecture matched to the metric volume.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_VEiIUkMU66dvdCPHHub8hg