Data Storage

Choose where records live, how they are keyed and queried, and how the store grows past a single machine. Storage is the hardest layer to change later: a wrong data model or shard key calcifies into a scaling ceiling, and getting replication wrong silently serves stale or lost data.

When to reach for this

Any system that persists state: picking SQL vs NoSQL, designing a schema and its access paths, adding indexes, splitting a hot table, distributing data across nodes (sharding/partitioning), adding read replicas, or deciding what to denormalize. Reach here the moment "store the data" needs a concrete key and query shape.

When NOT to

Don't shard, add replicas, or reach for NoSQL before a number forces it (YAGNI). A single well-indexed relational node handles ~1k QPS and tens of GB to low TB comfortably — most systems never outgrow it. Sharding multiplies operational cost and breaks joins/transactions; add it only when one node's write throughput or dataset size is genuinely exceeded (→ back-of-the-envelope). Caching reads (→ caching) and adding read replicas are cheaper first moves than sharding.

Clarify first

Answer these before choosing a store or topology — they decide the design:

• Data shape & relationships — flat key-value? rich relations needing joins?

document blobs? a graph of connections? Drives SQL vs NoSQL.

• Access patterns — how is data read and written, not just where it lives.

Point lookups by key, range scans, ad-hoc queries, aggregations? Model the store around the queries it must serve.

• Read:write ratio & scale — QPS each way, total size now and at retention.

(→ back-of-the-envelope for QPS, storage, and shard counts.)

• Consistency need — must reads see the latest write, or is eventual OK? Are

multi-record transactions required? (CAP/consistency theory → consistency-coordination.)

• Latency & durability targets — p99 read/write budget, and how much recent

data the system can afford to lose on a node failure.

The options

Relational (SQL — Postgres, MySQL): strict schema, joins, ACID transactions. Use when data is relational, integrity matters, and queries are varied/ad-hoc — the safe default until a number rules it out.

Document (MongoDB, etc.): flexible schema, self-contained JSON-ish documents queried by structure. Use when records are read/written as a whole and the schema evolves; relationships are few.

Key-value (Redis, DynamoDB, Riak): O(1) get/put by key, no rich queries. Use when access is purely by a known key and massive throughput is needed.

Wide-column (Cassandra, Bigtable, HBase): rows keyed by partition, columns sparse, keys kept sorted for range scans. Use when writes are huge and the table can be designed around a few known query patterns.

Graph (Neo4j): nodes and edges. Use when the core queries traverse many-to-many relationships (social graph, recommendations).

Scaling moves (apply on top of any store):

• Indexing — add a secondary structure so a query stops scanning. First lever.
• Read replicas (leader-follower) — copy writes to followers that serve reads.

Use when reads dominate and slight staleness is OK.

• Federation — split DBs by function (users / products / forums). Use when

functional domains scale independently and rarely join.

• Sharding/partitioning — split one logical table across nodes by a shard key.

Use when a single node's writes or dataset are exceeded. (This skill owns it.)

• Denormalization — store redundant copies to skip joins. Use when reads

vastly outnumber writes and joins are the bottleneck.

Polyglot persistence: use more than one of the above, each for what it's best at (e.g. Postgres for orders, Redis for sessions, a search index for full-text). The cost is operating and reconciling several stores.

Trade-offs

Behavior under stress

Storage is where load and failure get amplified into outages.

• Hot shard / celebrity key: a skewed shard key sends disproportionate traffic

to one node (a viral user, a single tenant) while others idle. The cluster looks under-loaded but one shard is melting. Mitigate: a shard key with high cardinality and even distribution; isolate or sub-partition hot keys.

• Replication lag: under write bursts, followers fall behind the leader, so

reads go stale — and a failover may promote a follower that lost recent writes. Mitigate: read-from-leader for read-your-writes; bound and alert on lag.

• Connection exhaustion: each DB connection costs memory and a backend

process/thread; a traffic spike (or a retry storm) opens more connections than the DB can serve, and every query slows or errors. A pooler in front (bounded pool) is what keeps the DB alive — without it the database is a SPOF that fails under load.

• Thundering writes / lock contention: hot rows or a single sequence/counter

serialize writes; index updates and lock waits stack up.

• Resharding under pressure: rebalancing while already overloaded moves huge

data volumes and can tip the cluster over. Plan capacity ahead of the cliff.

Monitor: replication lag, per-shard QPS and size (skew), connection-pool saturation/wait time, slow-query rate, lock waits, and disk/IOPS headroom.

How to apply

1. Clarify the inputs — settle data shape, access patterns, read:write ratio,

consistency need, and latency/durability targets (see Clarify first). No store choice survives unknown access patterns.

1. Pick the store from the trade-off table — match data shape to an option;

default to relational until a number or relationship pattern rules it out. Name what each candidate worsens, not just what it solves.

1. Pin the interface — write the primary key, partition/sort key, and the

secondary indexes for each query before adding scale machinery (see Interface sketch). The key is the decision.

1. Set the scaling knobs in cheap-first order — index, then cache (→ caching),

then read replicas, then federation, then shard. Stop at the first level that meets the target.

1. Stress-test the choice — walk hot shard, replication lag, connection

exhaustion, and resharding (see Behavior under stress); confirm a skewed key or write burst does not melt one node.

1. Size it, then pick a provider — compute shard count and per-node load from

back-of-the-envelope; if it is one node, do not shard. Default to the generic recipe and read the provider file only when a cloud is named.

Dos and don'ts

Do

• Start from the queries: model the store around its access paths, and write a

concrete key before drawing any box.

• Default to a single well-indexed relational node and exhaust index + cache +

replicas before sharding.

• Choose a shard key with high cardinality and even distribution; isolate or

sub-partition known hot keys.

• Put a bounded connection pooler in front of the database and alert on

replication lag and per-shard skew.

• Compute shard count from peak write QPS and dataset size, taking the larger.

Don't

• Don't reach for NoSQL, replicas, or shards before a number forces it (YAGNI).
• Don't pick a store on hype before the data shape and consistency need are known.
• Don't assume replicas give fresh reads — replication lag serves stale data and

failover can lose recent writes.

• Don't shard on a low-cardinality or monotonically increasing key; it creates hot

shards and write hotspots.

• Don't reshard a cluster that is already overloaded; plan capacity ahead of the

cliff.

Numbers that matter

Don't restate the tables — pull the figures from back-of-the-envelope. The ones that drive storage decisions: a single RDBMS node ≈ 1k QPS; a key-value node ≈ 10k QPS; 10 GB fits in RAM, 10 TB needs distributed storage. Use these to compute shard count = peak write QPS ÷ per-node QPS (and again by size = total bytes ÷ per-node capacity), then take the larger. If the result is one node, do not shard.

Interface sketch

The data model is the contract (GUIDE failure mode #8) — a "NoSQL box" decides nothing until the key is written. Pin down per entity:

• Primary key — what uniquely identifies a row/item and how it's looked up.
• Partition (shard) key + sort key — e.g. wide-column/DynamoDB:

PK = user_id (hash, spreads load), SK = created_at (sort, enables range scans like "latest N posts"). The PK must be high-cardinality and even.

• Secondary indexes — the non-key access paths that must be supported, each

with its query.

• Relational — tables, columns with types, foreign keys, and the indexes that

back each query; note what is intentionally denormalized.

Without a concrete key and the query it serves, the rest of the design is guesswork.

Choosing a provider

Default to the generic recipe above. If the user names a cloud, read references/providers/<provider>.md for the managed-service mapping, quotas/limits, and provider-specific trade-offs. If no file exists for that provider, the generic recipe is the answer.

Diagram

To visualize the data tier — leader with read replicas, a sharded cluster with a router, or a polyglot split — use the in-plugin architecture-diagram skill. Draw the shard key on the routing arrow and the replication direction explicitly; do not embed Mermaid here.

Related building blocks

• caching — pairs with this for read offload and is the cheap first move;

alternative to read replicas before sharding.

• consistency-coordination — owned-concept lives in it: CAP/consistency models,

consistent hashing, quorum, and distributed transactions/saga. Pairs with this when stale reads or cross-shard atomicity are unacceptable.

• back-of-the-envelope — feeds into this: supplies the QPS, storage, and

shard-count numbers that force (or rule out) each move here.

• scaling-evolution — depends on this block; sequences when each storage move

is introduced as a system grows.

• system-design — the orchestrator that routes into this block.

References

• `references/deep-dive.md` — partitioning schemes (range/hash/directory),

replication mechanics and conflict resolution, index internals (B-tree vs LSM), normalization vs denormalization, connection pooling, resharding. Read when designing the data tier in detail.

• `references/providers/{generic,aws,azure,gcp}.md` — service mappings, the

limits that change a decision, and per-environment pitfalls.