What causes a hot partition and how do you fix it?

Partitions have physical limits. Even if the cluster has spare capacity, one partition cannot use it. Common causes are monotonic keys such as timestamps, low-cardinality keys such as status, and celebrity or tenant skew where a few entities are far larger than the rest.

Detection

• Per-partition throughput and latency metrics, not just cluster averages
• Throttled requests concentrated on specific key ranges
• A single tenant or product with far more traffic than others
• Write amplification from a secondary index whose key is skewed

function saltedKey(entityId, bucketCount) {
  const bucket = Math.floor(Math.random() * bucketCount);
  return "ENTITY#" + entityId + "#" + bucket;
}

function bucketKeys(entityId, bucketCount) {
  const keys = [];
  for (let b = 0; b < bucketCount; b += 1) {
    keys.push("ENTITY#" + entityId + "#" + b);
  }
  return keys;
}

Write sharding helps writes but makes reads more complex, so use it only for genuinely hot keys. Another option is to cache the hot item in front of the database, or to split a large entity into smaller items that can be read independently. Each approach moves complexity somewhere else, so measure the trade-off.