Resources · 86

API pagination: complete exports without missing or duplicated records

Define ordering, tokens, authorization and recovery when a collection changes during traversal.

· 9 min

Global network map in an empty intelligence briefing room AI-generated editorial illustration

What this guide helps achieve

  • Define ordering and scope
  • Treat tokens as continuation
  • Detect completion from the contract
  • Recover after failures
  • Reconcile the result

Quick check

  • Does an empty page mean completion?
  • Can a cursor provide authorization?
  • Does deduplication prevent missing records?

Step-by-step method

  1. 01

    Define ordering and scope

    Document filters, ordering and a tie-breaking key. A timestamp alone may be insufficient when several objects share it. Decide whether the collection stays live or the export uses a snapshot. Do not promise a frozen view when the provider offers no such guarantee.

    Deliverable: ordering and consistency contract.

  2. 02

    Treat tokens as continuation

    AIP-158 specifies opaque tokens and separates pagination from authorization: a token does not confer access rights. Check identity and permissions on every request. Keep reading parameters consistent with the contract and handle expiry or invalidation explicitly.

    Deliverable: valid, expired and unauthorized token cases.

  3. 03

    Detect completion from the contract

    Do not assume completion because a page is shorter than requested. An API may return fewer items while providing continuation. Under AIP-158 an empty next_page_token indicates completion; adapt clients to the provider’s actual contract. Also detect repeated tokens without progress.

    Deliverable: completion and loop tests.

  4. 04

    Recover after failures

    Save continuation after durable storage of the processed batch. Deduplicate using stable identifiers and keep recovery state separate from displayed totals. If a snapshot expires, restart or reconcile under an explicit policy; do not silently concatenate two different collection states.

    Deliverable: checkpoint and failure scenario.

  5. 05

    Reconcile the result

    Compare unique identifiers, exclusions, states and any reference total at a common cutoff. Test creation, changes and deletion during reading. A correct count does not prove the right objects were exported. Protect files and traces according to their data.

    Deliverable: export manifest and documented differences.

A paginated read that can recover

Under an AIP-158-style contract, continuation determines the next step. A short page does not establish completion.

  1. Define the read

    Retain filter, order, parameters and authorized identity.

  2. Store the batch

    Durably save objects and their identifiers.

  3. Follow continuation

    A present token leads to the next request with bound parameters.

  4. Reconcile at completion

    When the final token is empty, check exclusions, reference state and coverage.

Fictional example: the connection fails after batch storage. Recovery uses the saved checkpoint and checks identifiers already processed.

Right count, wrong objects: trace a deletion between pages

Fictional offset API example with a page size of 3 and alphabetical identifier order. The collection stays live; no snapshot is promised. These identifiers and responses are teaching data.

StageCollection at that timeRequest or resultWhat the test establishes
Initial stateA, B, C, D, E, Foffset=0, limit=3 → A, B, CThe client stores A, B and C. Its next planned position is 3.
B is deletedA, C, D, E, FThe collection now contains 5 objects.D moves from the fourth to the third position; an offset is a position, not an identity.
Second pageA, C, D, E, Foffset=3, limit=3 → E, FD is never read. The client has stored A, B, C, E, F in total.
Final reconciliationFinal reference: A, C, D, E, F5 received identifiers; 5 expected identifiersCounts agree, but D is missing and B is extra relative to the final state. Comparing identities exposes the difference.

Acceptance case to reproduce

Fictional example: these inputs describe no customer or observed result.

View case inputs
{
    "page_1_ids": [
        "A",
        "B"
    ],
    "page_2_ids": [
        "B",
        "C"
    ],
    "next_cursor": "c2",
    "empty_page_has_next_cursor": true
}

Expected decision

B repeats. Check identities and continue while a next cursor exists, even after an empty page.

Verification procedure

  1. Fix filters, stable order and object identity.
  2. Test duplicates and an empty page with a next cursor.
  3. Keep identifiers, cursors and the actual end condition. Reconcile received objects and unique identifiers; flag repeated cursors without treating an empty page with a continuation as collection completion.

Google AIP-158 — Pagination

Reusable worksheet

Complete with your authorised observations. These fields are a working template, not observed results.

FieldInformation to record
CollectionFilters, order, tie-breaker and guarantees
ContinuationToken, linked parameters, expiry and completion
RecoveryDurable batch, unique identifiers and saved checkpoint
ResultManifest, exclusions, cutoff and reconciliation differences

Worked example

Illustrative situation

Deduplication changes nothing in the fictional case above: all five received identifiers are already distinct. The decisive check is reconciliation against a reference state with an explicit date and scope.

Decision and expected evidence

If the task requires a fixed state, use a snapshot guarantee actually provided by the supplier. A cursor alone does not establish one. For a live read, document what reconciliation verifies and what remains unknown; do not present a matching count as proof of completeness.

Distinguish the mechanisms

MechanismPurposeCheck or limitation
OffsetAccess a numeric positionInsertions can shift positions
CursorContinue from a reading markerDoes not itself guarantee a snapshot
SnapshotFix a reference stateCheck validity period and cost

Management indicators

IndicatorWhat it measuresFirst action
Unique identifiersDistinct objects retained in the exportCompare with an equivalent reference scope
ProgressTokens and batches genuinely advancingStop loops and investigate
Verified coverageObjects reconciled with a reference stateDo not equate volume with completeness

Common pitfalls

  • Insertions can shift positions
  • Does not itself guarantee a snapshot
  • Check validity period and cost

Frequently asked questions

Does an empty page mean completion?

Check the contract and continuation token. Page size alone should not determine completion.

Can a cursor provide authorization?

No. The server must check request permissions independently of the pagination token.

Does deduplication prevent missing records?

No. It removes repeats without recovering unread objects. Test ordering stability and collection changes.

Pagination: consistency and recovery

Reaching every page does not prove a coherent collection

A paginated collection needs an object order, continuation rules and a reconciliation reference. An advancing cursor can miss a row whose key changes; a correct unique-ID count can mix revisions. This exercise makes those differences checkable on five fictional objects.

Offline simulation: no API request, database or export file is created. Objects A–E, TRAIN tokens and snapshot S1 are teaching fixtures. Results describe this model and guarantee neither provider behavior nor coverage of a real collection.

Initial data: three objects share one timestamp

Initial data: three objects share one timestamp
Identifierupdated_at in UTCRevision
A2026-10-22T10:00:00Z1
B2026-10-22T10:00:00Z1
C2026-10-22T10:00:00Z1
D2026-10-22T10:01:00Z1
E2026-10-22T10:02:00Z1

Ascending (updated_at, id), maximum page size 2, unique A–E IDs and non-null sort fields. Dates use equal precision. Production contracts also need types, collation and NULL handling; the word “cursor” supplies none of those assumptions.

Five computed reads and their references

These five runs model query-boundary reads: a query returning no rows ends each local run. A remote API’s completion rule is a separate contract; the continuation chain below does not stop at an empty page.

P1 · Timestamp alone leaves C behind

After A and B, updated_at > 10:00 skips C because it shares their timestamp. The ordering includes an ID tie-breaker, but the continuation condition omits it.

Reconciliation reference: Unchanged initial dataset

P1 · Timestamp alone leaves C behind
Page and collection readReceived objectsPost-page boundary
1 · Live collection at this readA/r1 · 2026-10-22T10:00:00Z ; B/r1 · 2026-10-22T10:00:00Z2026-10-22T10:00:00Z
2 · Live collection at this readD/r1 · 2026-10-22T10:01:00Z ; E/r1 · 2026-10-22T10:02:00Z2026-10-22T10:02:00Z
3 · Live collection at this readNoneEnd observed in the model
Received rows
4
Unique identifiers
4
Identifiers missing from the reference
C
Repeated identifiers
None
Identifiers received under multiple revisions
None

Four rows arrive and C is missing. A correct batch size does not repair an incomplete boundary.

P2 · Timestamp/ID pair on a stable collection

The boundary retains the last object’s timestamp and ID. After (10:00, B), C remains greater in this order. All five objects are read once in this stable dataset.

Reconciliation reference: Unchanged initial dataset

P2 · Timestamp/ID pair on a stable collection
Page and collection readReceived objectsPost-page boundary
1 · Live collection at this readA/r1 · 2026-10-22T10:00:00Z ; B/r1 · 2026-10-22T10:00:00Z2026-10-22T10:00:00Z / B
2 · Live collection at this readC/r1 · 2026-10-22T10:00:00Z ; D/r1 · 2026-10-22T10:01:00Z2026-10-22T10:01:00Z / D
3 · Live collection at this readE/r1 · 2026-10-22T10:02:00Z2026-10-22T10:02:00Z / E
4 · Live collection at this readNoneEnd observed in the model
Received rows
5
Unique identifiers
5
Identifiers missing from the reference
None
Repeated identifiers
None
Identifiers received under multiple revisions
None

This result depends on an unchanged collection and total order. It promises no frozen view for a live API.

P3 · A moves forward and returns with a new revision

After page one, A changes from 10:00/r1 to 10:03/r2. The pair cursor advances but encounters A again.

Reconciliation reference: Final live collection

P3 · A moves forward and returns with a new revision
Page and collection readReceived objectsPost-page boundary
1 · Live collection at this readA/r1 · 2026-10-22T10:00:00Z ; B/r1 · 2026-10-22T10:00:00Z2026-10-22T10:00:00Z / B
2 · Live collection at this readC/r1 · 2026-10-22T10:00:00Z ; D/r1 · 2026-10-22T10:01:00Z2026-10-22T10:01:00Z / D
3 · Live collection at this readE/r1 · 2026-10-22T10:02:00Z ; A/r2 · 2026-10-22T10:03:00Z2026-10-22T10:03:00Z / A
4 · Live collection at this readNoneEnd observed in the model
Received rows
6
Unique identifiers
5
Identifiers missing from the reference
None
Repeated identifiers
A
Identifiers received under multiple revisions
A

Six rows and five unique IDs arrive. A occurs as revisions 1 and 2. ID deduplication alone does not establish temporal consistency of values; revision handling needs an explicit policy.

P4 · C moves behind the cursor

After page one, C changes from 10:00/r1 to 09:59/r2. Its new key falls below the boundary already passed.

Reconciliation reference: Final live collection

P4 · C moves behind the cursor
Page and collection readReceived objectsPost-page boundary
1 · Live collection at this readA/r1 · 2026-10-22T10:00:00Z ; B/r1 · 2026-10-22T10:00:00Z2026-10-22T10:00:00Z / B
2 · Live collection at this readD/r1 · 2026-10-22T10:01:00Z ; E/r1 · 2026-10-22T10:02:00Z2026-10-22T10:02:00Z / E
3 · Live collection at this readNoneEnd observed in the model
Received rows
4
Unique identifiers
4
Identifiers missing from the reference
C
Repeated identifiers
None
Identifiers received under multiple revisions
None

C is missing despite the complete sorting pair. A mutable sort key needs a consistency strategy beyond a unique tie-breaker.

P5 · Reading S1 retains S1 identities and values

A moves forward and C backward in the live collection, but every page in this run reads the exercise’s immutable S1 copy.

Reconciliation reference: Fictional snapshot S1

P5 · Reading S1 retains S1 identities and values
Page and collection readReceived objectsPost-page boundary
1 · Fictional snapshot S1A/r1 · 2026-10-22T10:00:00Z ; B/r1 · 2026-10-22T10:00:00Z2026-10-22T10:00:00Z / B
2 · Fictional snapshot S1C/r1 · 2026-10-22T10:00:00Z ; D/r1 · 2026-10-22T10:01:00Z2026-10-22T10:01:00Z / D
3 · Fictional snapshot S1E/r1 · 2026-10-22T10:02:00Z2026-10-22T10:02:00Z / E
4 · Fictional snapshot S1NoneEnd observed in the model
Received rows
5
Unique identifiers
5
Identifiers missing from the reference
None
Repeated identifiers
None
Identifiers received under multiple revisions
None

All five revision-1 objects arrive once. S1 is the reconciliation reference; the final live state can differ. An opaque token alone does not establish that a provider implements this snapshot.

A snapshot needs a contract separate from the cursor

A sorting pair defines a boundary; a snapshot defines the objects and values visible during reading. PostgreSQL 18 distinguishes per-command Read Committed views from a Repeatable Read transaction view. An API spanning requests needs to explain its own mechanism and availability period. This exercise opens no transaction and does not recommend indefinitely retaining a production transaction.

An empty page can still have a next page

Under AIP-158, a response can return fewer objects than requested, including zero, without being final. An empty next_page_token signals completion. The TRAIN labels below are public teaching fixtures; clients must not infer a numeric position from them.

An empty page can still have a next page
ResponseSent tokenObjectsReturned continuation
L1StartA, BTRAIN-CONT-1
L2TRAIN-CONT-1NoneTRAIN-CONT-2
L3TRAIN-CONT-2CTRAIN-CONT-3
L4TRAIN-CONT-3D, EEnd observed in the model

Stop a repeated token without claiming completion

Stop a repeated token without claiming completion
ResponseSent tokenObjectsReturned continuation
X1StartA, BTRAIN-LOOP
X2TRAIN-LOOPNoneTRAIN-LOOP

Stop: repeated continuation

The second X response repeats the token already used. The teaching client stops this lack of progress. A real API client also needs bounds on calls, time, memory and retries: a changed token alone does not establish useful progress.

Nine continuation requests: context and permissions

The model binds a token to parent, filter, order and S1. Page size can change under AIP-158’s explicit exception. server_authorized and snapshot_available are server-side facts supplied by the fixture, not authoritative client parameters. Fictional expiry is 22 October 2026 at 13:00 UTC, an exclusive boundary.

C01 · Same context and current permission
{
    "token": "TRAIN-CONT-1",
    "parent": "training/catalogue-a",
    "filter": "state=active",
    "order": "updated_at,id",
    "page_size": 2,
    "now": "2026-10-22T12:00:00Z",
    "server_authorized": true,
    "snapshot_available": true
}

Continuation admissible in the profile

Continuation is admissible in this profile; permission checks still apply to returned objects.

C02 · Only page size changes
{
    "token": "TRAIN-CONT-1",
    "parent": "training/catalogue-a",
    "filter": "state=active",
    "order": "updated_at,id",
    "page_size": 1,
    "now": "2026-10-22T12:00:00Z",
    "server_authorized": true,
    "snapshot_available": true
}

Continuation admissible in the profile

Changing 2 to 1 does not by itself break this pagination context. Do not assume a provider with another contract follows the same rule.

C03 · Different parent
{
    "token": "TRAIN-CONT-1",
    "parent": "training/catalogue-b",
    "filter": "state=active",
    "order": "updated_at,id",
    "page_size": 2,
    "now": "2026-10-22T12:00:00Z",
    "server_authorized": true,
    "snapshot_available": true
}

Reject: different read context

A token must not become a bridge into a different collection parent.

C04 · Different filter
{
    "token": "TRAIN-CONT-1",
    "parent": "training/catalogue-a",
    "filter": "state=archived",
    "order": "updated_at,id",
    "page_size": 2,
    "now": "2026-10-22T12:00:00Z",
    "server_authorized": true,
    "snapshot_available": true
}

Reject: different read context

Reusing a token after changing filters would mix two different reads.

C05 · Different order
{
    "token": "TRAIN-CONT-1",
    "parent": "training/catalogue-a",
    "filter": "state=active",
    "order": "id",
    "page_size": 2,
    "now": "2026-10-22T12:00:00Z",
    "server_authorized": true,
    "snapshot_available": true
}

Reject: different read context

A boundary created for one order must not silently acquire a different interpretation.

C06 · Permission revoked
{
    "token": "TRAIN-CONT-1",
    "parent": "training/catalogue-a",
    "filter": "state=active",
    "order": "updated_at,id",
    "page_size": 2,
    "now": "2026-10-22T12:00:00Z",
    "server_authorized": false,
    "snapshot_available": true
}

Reject: current permission absent

The token still exists but grants no permission. Earlier authorization does not authorize the current request.

C07 · Exact expiry instant
{
    "token": "TRAIN-CONT-1",
    "parent": "training/catalogue-a",
    "filter": "state=active",
    "order": "updated_at,id",
    "page_size": 2,
    "now": "2026-10-22T13:00:00Z",
    "server_authorized": true,
    "snapshot_available": true
}

Recover under policy: token expired

At 13:00:00Z the token is expired under this profile. Its lifetime is fictional and supplies no universal retention recommendation.

C08 · Snapshot unavailable
{
    "token": "TRAIN-CONT-1",
    "parent": "training/catalogue-a",
    "filter": "state=active",
    "order": "updated_at,id",
    "page_size": 2,
    "now": "2026-10-22T12:00:00Z",
    "server_authorized": true,
    "snapshot_available": false
}

Review: S1 reference unavailable

The profile refuses to silently concatenate S1 with a new view. Restart or reconcile under an explicit policy.

C09 · Unknown token
{
    "token": "TRAIN-UNKNOWN",
    "parent": "training/catalogue-a",
    "filter": "state=active",
    "order": "updated_at,id",
    "page_size": 2,
    "now": "2026-10-22T12:00:00Z",
    "server_authorized": true,
    "snapshot_available": true
}

Reject: unknown token

The model accepts only tokens in its fictional registry; it does not decode received text and invent continuation semantics.

Two crash recoveries: the checkpoint changes coverage

S1 batches are A/B, C/D and E. In this in-memory model, “durable” represents a batch already confirmed in fictional local storage. This is no real persistence test. Local processing is idempotent by S1 ID; that does not guarantee an exactly-once external effect.

Before or after the batch: two consequences
Before or after the batch: two consequences Advancing the checkpoint before a batch leaves A and B missing after a crash. Retaining the batch before checkpointing allows rereading and deduplication, recovering all five S1 IDs. Checkpoint before the batch Crash before retaining A/B A/B skipped on recovery Missing: A, B Batch before the checkpoint Crash before checkpointing A/B reread and deduplicated Missing: None
R1 · Checkpoint before retaining the batch

R1 records the next page before retaining A/B. A crash occurs between the operations; recovery skips A and B.

R1 · Checkpoint before retaining the batch
StepSimulated operationRetained objectsRecorded next page
1Record the next page · 2None2
2Crash; volatile memory lostNone2
3Resume at the recorded pageNone2
4Retain the batch idempotently · 2C, D2
5Record the next page · 3C, D3
6Retain the batch idempotently · 3C, D, E3
7Record the next page · 4C, D, EEnd observed in the model

Identifiers missing from the reference: A, B

R2 · Batch retained before checkpoint

R2 retains A/B, then crashes before checkpointing. Recovery rereads A/B; the idempotent sink retains them once, then finishes S1.

R2 · Batch retained before checkpoint
StepSimulated operationRetained objectsRecorded next page
1Retain the batch idempotently · 1A, B1
2Crash; volatile memory lostA, B1
3Resume at the recorded pageA, B1
4Retain the batch idempotently · 1A, B1
5Record the next page · 2A, B2
6Retain the batch idempotently · 2A, B, C, D2
7Record the next page · 3A, B, C, D3
8Retain the batch idempotently · 3A, B, C, D, E3
9Record the next page · 4A, B, C, D, EEnd observed in the model

Identifiers missing from the reference: None

When batch and checkpoint share a suitable local transaction, test their joint commit. Effects across systems need a recovery protocol and reconciliation; simply reordering two calls does not demonstrate distributed atomicity.

Six records for pagination acceptance checks

  • Scope, filter, total order and rules for nullable or mutable keys.
  • Consistency reference: live collection, identified snapshot or another documented contract.
  • Received batches, IDs and revisions, with reconciliation differences.
  • Handling of short/empty pages, repeated tokens and bounded resources.
  • Confirmed recovery checkpoint, local deduplication and completed external effects.
  • Current permission checks and handling of unavailable tokens or snapshots.

The five reads, nine requests and two recoveries are computed editorial counterexamples. AIP-158 applies only to the declared contract; actual guarantees must come from the provider. No client export, real completeness rate or database transaction is claimed.

Primary sources and scope

Sources reviewed on

Official references

References consulted: . The method and worksheet propose checks to adapt to your context; they do not constitute certification.