NoSQL Databases

When NoSQL fits

Managed NoSQL databases scale to huge request rates with predictable latency, as long as you design for your access patterns up front.

Partition keys

Data is spread across partitions by a partition key. Every partition has throughput limits (in DynamoDB, about 3,000 reads and 1,000 writes per second), so a key that receives a large share of traffic becomes a hot partition and gets throttled even if the table as a whole has spare capacity.

import hashlib
from collections import Counter

PARTITIONS = 4
def partition(key):
    return int(hashlib.md5(key.encode()).hexdigest(), 16) % PARTITIONS

writes_by_date = Counter(partition("2026-11-27") for _ in range(4000))           # key = today's date
writes_by_order = Counter(partition(f"order-{i}") for i in range(4000))         # key = order id
print("key=date:     writes per partition", [writes_by_date.get(p, 0) for p in range(PARTITIONS)])
print("key=order_id: writes per partition", [writes_by_order.get(p, 0) for p in range(PARTITIONS)])

key=date:     writes per partition [0, 0, 4000, 0]
key=order_id: writes per partition [986, 1019, 972, 1023]

Good keys have many distinct values with evenly spread traffic: customer_id, order_id, or a key with a suffix spread across N shards for very hot items.