CAPACITY NOTES
How often to take a full snapshot audit
Delta polling trusts the chain of every past state. A weekly full read per sport is 52 calls a year and settles the doubt permanently.
Delta polling keeps your store in step with the feed by reading only what changed, and it works so well that a quiet doubt creeps in: what if a change was missed? A drifted cursor, a bug in the merge, a week of silent skew. The answer is not more deltas. It is the occasional full read, scheduled like an audit.
Deltas trust the chain
Every since read assumes the last state was right. If one delta was applied wrong, every state after it carries the error forward, and nothing in the next delta contradicts it. That is the shape of the risk: not losing the cursor, which is loud, but corrupting state quietly while the cursor advances happily, as the cursor recovery post on this site separates the two failures.
A full read is the only comparison that settles it. One board, every event, exactly as the source sees it now.
The audit is cheap
The numbers are small on purpose. The REST polling guide documents the contrast: a full prematch soccer board can carry 1,522 events against a handful for a delta, checked 2026-11-04. Once a week per sport, that contrast costs one request and one heavier response. A year of weekly audits per sport is 52 calls: nothing, against a daily delta budget.
What you get for those 52 calls is proof that the store equals the source, or a diff to chase. Either way, the doubt is priced out of the system.
What to compare
Compare identity and price, not payload. For each event the audit read returns, check that your store holds it, holds it with the same key, and holds the same current prices for the markets you track. Events the audit shows that you lack are missed inserts. Events you hold that the audit no longer shows are stale rows, left behind by matches that started or markets that closed.
Diffs are rarely zero on a live board, because prices move between your last delta and the audit read. Compare against your last delta's timestamp, and treat a diff as real only when it survives a second read.
When the diff says something
A recurring diff of the same kind is a bug with a signature. Events you keep missing point at the merge. Rows that never leave point at the cleanup. Prices that lag by exactly one update point at the cursor ordering, applied after the write instead of before it.
Fix the class, not the row. The audit's value is the pattern, and the pattern only shows up if the audit runs on a schedule instead of on suspicion.
Run the deltas for freshness, the audit for truth. A store that is verified weekly is a store you stop worrying about.