Architecture & System Design › Cloud Design Patterns · also in Queues & Async Processing

Claim Check

Putting a large payload in storage and sending only a reference through the queue.

Also known as: claim check, claim-check pattern, reference passing

The claim check pattern passes large payloads by reference: store the data once (object storage, database, cache), send only its key (“claim ticket”) through queues and events, and let consumers fetch when ready. Messages stay small and fast; bulk data moves out-of-band exactly once.

producer → store blob → emit {claimCheck: "s3://…/uuid"} (tiny)
consumer → fetch blob → process → optionally delete

It suits large documents, images, batch files and ML payloads flowing through size-limited transports (message brokers cap messages in KB–MB; databases dislike blobs in hot rows). Small control plane, big data plane — each optimised separately.

The classic mistakes:

• Blobs in messages. Stuffing megabytes into queues breaks size limits, slows brokers, and duplicates storage per subscriber. Reference, don’t embed.
• Orphaned blobs. Produced payloads whose messages never arrive (or consumers crash pre-cleanup) accumulate forever. TTLs plus reaper jobs bound the orphans.
• No integrity check. Fetching corrupted or partial blobs processes garbage authoritatively. Checksum at store; verify at fetch.
• Premature deletion. Producers cleaning up before slow consumers fetch loses data silently. Delete after acknowledgement (or TTL), never on publish.
• Security by obscurity. Unguessable URLs treated as authorisation leak to anyone obtaining them (logs, forwards). Presign narrowly, expire quickly, authorise fetches.
• Hot-spot fetches. Thousands of consumers fetching one blob simultaneously stampede storage; CDN/cache the hot payloads.
• Version skew. Blob overwritten while consumers fetch mixes generations. Immutable blobs (new key per version), never mutate-in-place.

When to use it: payloads exceeding comfortable message sizes, or consumed selectively (not every subscriber needs the bytes). Small messages, referenced bulk — each transport doing what it does best.