* feat(events): fetch can start from the newest event GET /events/fetch takes from=end: with no cursor it answers an empty page whose cursor sits at the newest row of the slice, so later reads return only rows created or replaced after it. puter.events.fetch() passes `from` through. Adds a (user_id, audience, app_uid, id) index on notification so the newest row of a slice is one seek. * fix(puter-js): onPush starts its feed from the newest notification A notifications listener started without catch-up asks for from: 'end' and keeps the cursor it gets back instead of reading every stored notification to mark them as seen. A server that ignores `from` answers with the oldest page, and the listener reads through it as before. * fix(events): read the newest notification id from the primary * fix(events): mark the end cursor at the slice's latest replacement The end cursor's replacement mark is the newest (replaced_at, id) in the slice rather than the serving node's clock, since replaced_at is stamped by whichever node wrote it. * docs(events): describe the end cursor's replacement mark
5.1 KiB
Pagination convention for list APIs
Every list endpoint follows one wire contract. This document is the source of truth for adding pagination to a new or existing list surface.
Wire contract
Requests accept:
| Param | Type | Meaning |
|---|---|---|
limit |
number | Maximum items per page. Each endpoint documents its cap and default. |
cursor |
string | null | Opaque continuation token. null (or any presence of the key) requests the first page. |
offset |
number | Skip N items. Legacy/discouraged — see below. Cannot be combined with cursor. |
includeTotal |
boolean | Adds total to the response. |
Paginated responses are an envelope:
{ "items": [...], "cursor": "…", "total": 123 }
cursoris present only while more pages exist. Clients iterate until it is absent.totalis present only when the request setincludeTotal.- Pages may be short. Post-query filtering (TTL expiry, permission
checks) can shrink a page below
limit— or even to zero — while acursoris still returned. Never useitems.length < limitas an end-of-list signal.
Backward compatibility
Requests without pagination params keep returning the full result in the legacy shape (bare array) forever — old clients never break.
The envelope trigger depends on the endpoint's history:
- Endpoints where
limitpre-dates the convention (readdir, subdomainselect) return the envelope only when the request containscursor(includingnull) orincludeTotal;limit/offset-only requests keep the bare array. - Endpoints where every pagination param is new (kv
list, workersgetFilePaths) return the envelope when any pagination param is present.
Cursors
Cursors are opaque base64-encoded JSON, produced and consumed only by the
backend (src/backend/util/pagination.ts). A cursor over a global sequence
(an auto-increment id) is sealed with sealCursor/openCursor instead, so
two pages can't be decoded into how fast the table grows; openCursor still
reads a plain cursor. What a cursor wraps is an implementation detail per
store:
- DynamoDB-backed lists wrap
LastEvaluatedKey. A kv listing of another namespace (cross-app or through a share handle) resumes after the last entry it returned instead, since the scan can stop on a private one, and its cursor is sealed to that listing and padded to one length. - SQL-backed lists wrap a keyset position —
(sortValue, id)of the last row — and the query seeks past it (WHERE (col, id) > (?, ?) ORDER BY col, id). Any SQL list gaining a cursor must have a deterministicORDER BYending in a unique tiebreaker (id). - The notification catch-up (
GET /events/fetch) holds two keysets in one sealed cursor: the lastidit returned, and the last(replaced_at, id)among rows up to that id. Each is its own indexed read, merged into one page, so a row replaced after a client read it comes back from that cursor.from=endwith no cursor answers an empty page whose cursor holds the slice's newestidand the slice's latest(replaced_at, id)as the replacement mark.
Cursors that carry a sort also pin it: a request that passes a cursor plus a conflicting sort is rejected with 400.
Offset
SQL-backed endpoints support offset natively. DynamoDB-backed endpoints
(kv) emulate it by advancing past skipped items with Select: COUNT queries
— no item data is transferred, but read capacity is still consumed for
everything skipped, and the caller is metered for it. Offset is supported
for parity, not recommended — cost grows linearly with the offset, so it is
capped (kv: 5000). Use cursors.
Totals (includeTotal)
- SQL:
SELECT COUNT(*)with the same WHERE clause as the listing. - DynamoDB: a
Select: COUNTloop over the query (TTL-filtered), metered to the caller. Cost is proportional to the total item count — request the total on the first page only, not every page. - Where visibility is decided per-actor after the query (protected apps in
the catalog listing),
totalapproximates the visible set: it counts non-protected plus caller-owned rows, and misses rows visible only through explicit permission grants.
fetchUntilFull (kv only)
DynamoDB applies Limit before its filter expression, so TTL-filtered pages
are structurally short. fetchUntilFull: true makes the backend keep
fetching (bounded number of continuation queries) until the page holds
limit items or the keyset is exhausted. Requires limit. If the bound is
hit, the response simply carries a cursor — still convention-legal.
Adding pagination to a new endpoint
- Use
encodeCursor/decodeCursor/normalizeLimit/normalizeOffsetfromsrc/backend/util/pagination.ts—sealCursor/openCursorwhen the cursor wraps a global id. - Push equality filters into the query so pages and counts operate on the true result set; only genuinely per-actor filtering may remain post-query (short-pages rule covers it).
- Fetch
limit + 1rows to detect whether a next page exists (SQL), or useLastEvaluatedKey(DynamoDB). - Keep the no-params request returning the legacy full result.