Aerospike Data Modeling

Aerospike: data model design

Role

Act as a data architect. Your job is to turn requirements into a durable schema contract, not to write client code. The output is documents that a team reviews and implements against.

When to use this skill

Use it when the starting point is requirements without a schema:

• A new application or service with no Aerospike model yet.
• A redesign, where an existing model no longer serves its access patterns.
• A review of a proposed schema for structural defects before implementation.

Not this skill: writing or reviewing client code against an existing model, tuning policies, choosing CDT operations, debugging a slow batch read. That is aerospike-development. If a schema already exists and the question is "how do I use it well," hand off.

What you produce

Two documents. Write them to files — they are review artifacts with a life beyond the session, not chat output.

The schema summary is generated from the schema guide, never authored independently. If they disagree, the schema guide wins and the summary is regenerated.

See references/model-deliverables-schema-guide-summary.md.

Mental model for data architects

Aerospike is neither a relational database nor a document database.

• Records are semi-structured. A record is a collection of strongly typed bins, and the typing is per bin per record — there is no set-level schema. Two records in the same set can have entirely different bins, and the server enforces nothing. Absent bins cost nothing, so sparse and heterogeneous shapes are cheap rather than wasteful. The consequence for design: the data model is an application-level contract — namespace, set, key format, bin names, and bin types that every client agrees on — and nothing in the database will stop a client that writes a different shape. Write the contract down; that is what the schema guide is for.
• Records are the unit of I/O. Record data is stored contiguously, so every read fetches the entire record from storage, and every write rewrites the entire record — Aerospike does not do in-place updates. Requesting a subset of bins trims what crosses the network, not what is read from device. A record in the tens of KiB therefore spends tens of KiB of I/O on every access, no matter how small the change. Record size is an I/O budget, not just a storage number.
• There are no server-side joins. The multi-record tool is the batch read, which scatters and gathers across nodes in parallel.
• Every record costs 64 bytes of primary index metadata, per replica, usually in RAM. Many tiny records spend more memory on index than on data.
• Access patterns drive the model — not entity normalization, and not document embedding.
• Consolidate, but bound it. Enough to avoid tiny records; not so much that one record becomes a monolith or a hot key.

If your instinct is a table per entity and a row per sub-entity, or one giant embedded document, you will produce a bad Aerospike model.

Record sizing — target band, the configured max-record-size limit, and the architectural ceiling are three different bounds that are easy to conflate. Do not carry a number from memory; read the current values from the data modeling guide (see Escalation below).

Whatever the band's endpoints are, read it as a distribution, not a target: design so the bulk of records sit at the low end (single-digit KiB), and treat the upper end as headroom for outliers and slowly-changing consolidated structures — 1:N and N:M relationship lists, where one record per edge would cost more. Size only hurts once multiplied by write frequency: a large record on a hot write path is a design defect even when it fits, because every update rewrites it in full — but the same size where writes are infrequent relative to reads is a legitimate design, not a compromise. Ask for the update rate, not just the byte count.

Do not design without clarifying first

The first deliverable is a written clarification document, not a schema. Ask requirements-gap questions — "what is the p95 fan-out?", "is eventual consistency acceptable here?" — never mechanism-preference questions like "which pattern do you prefer?". If deterministic guidance already resolves a choice, apply it instead of asking.

Do not fill gaps with assumptions and continue. When entity ownership, lifecycle, cardinality, or an access path is unclear, stop and ask. Where an input cannot be obtained, record it as an explicit assumption with a reconsider trigger rather than burying it.

Design one entity group at a time and pass its review before starting the next. See references/model-design-time-workflow.md.

Failure modes to check while drafting

Seven ways Aerospike models go wrong. Check them during design, not after. Each has a detection test in references/model-failure-modes-checklist.md.

1. Record granularity comes from cardinality and who drives the read — never from the entity list. One set per domain noun means the model came from an ER diagram.
2. The most frequent reads must be key lookups or bounded batch reads. If more than one or two access patterns resolve via secondary-index query, fix the keys, not the indexes.
3. Single-element mutations happen server-side, in place. Any read-modify-write of a whole bin should have been a CDT operation.
4. A bin is a container, not a field. Bin counts that scale with data rather than schema belong in one CDT bin. Bin names cap at 15 characters.
5. Duplicate data deliberately when two access patterns need it in two shapes. A second round trip purely to assemble a response is a normalization you should have collapsed.
6. Every collection bin needs a growth ceiling and a decided behavior at it. If element count is driven by user behavior rather than a design decision, it is unbounded.
7. Small independent entities still need an explicit sizing decision. Index overhead against a small payload is real cost; consolidating all of them into one record is the opposite error.

Escalation: use the data modeling guide

This skill covers the decision layer. The full workflow — the clarification gates, the per-relationship decision packs, the sizing worksheets, the stakeholder checkpoints — lives in the https://github.com/aerospike/data-modeling-guide repository.

For a new application, fetch the guide and follow its checklist. Do not design a complete model from this skill alone.

gh repo clone aerospike/data-modeling-guide

Then read its AGENTS.md first — it carries the routing table and the hard rules — followed by new-app-modeling-checklist.md.

If you cannot reach the guide — no access, no gh auth, offline — say so plainly and state what that limits. Deliver what this skill supports (the mental model, the failure-mode checks, a clarification document) and flag that the sizing worksheets and decision packs were not applied. Do not improvise a complete model and present it as if the full process ran.

Version-gated features

Several patterns depend on server version. Confirm the target version and client support before recommending any of them; the guide's checklist has a version gate table with fallbacks for each.

• Path expressions — nested CDT filtering and indexing.
• Expression indexes — sparse or computed-value indexing.
• Multi-record transactions — atomic multi-record updates; require a strong-consistency namespace, and carry limits that rule them out for wide cascades. See concepts-and-patterns.md § Multi-record consistency.

Do not state a specific minimum version from memory. Read it from the guide or the client matrix.

See also

• reference.md — external links.
• aerospike-development — implementation-time work against an existing model.