incremental-fetch

Incremental Fetch

Build data pipelines that never lose progress and never re-fetch existing data.

The Two Watermarks Pattern

Track TWO cursors to support both forward and backward fetching:

Watermark	Purpose	API Parameter
newest_id	Fetch new data since last run	since_id
oldest_id	Backfill older data	until_id

A single watermark only fetches forward. Two watermarks enable:

• Regular runs: fetch NEW data (since newest_id)
• Backfill runs: fetch OLD data (until oldest_id)
• No overlap, no gaps

Critical: Data vs Watermark Saving

These are different operations with different timing:

What	When to Save	Why
Data records	After EACH page	Resilience: interrupted on page 47? Keep 46 pages
Watermarks	ONCE at end of run	Correctness: only commit progress after full success

fetch page 1 → save records → fetch page 2 → save records → ... → update watermarks

Workflow Decision Tree

First run (no watermarks)?
├── YES → Full fetch (no since_id, no until_id)
└── NO → Backfill flag set?
    ├── YES → Backfill mode (until_id = oldest_id)
    └── NO → Update mode (since_id = newest_id)

Implementation Checklist

1. Database: Create ingestion_state table (see patterns.md)
2. Fetch loop: Insert records immediately after each API page
3. Watermark tracking: Track newest/oldest IDs seen in this run
4. Watermark update: Save watermarks ONCE at end of successful run
5. Retry: Exponential backoff with jitter
6. Rate limits: Wait for reset or skip and record for next run

Pagination Types

This pattern works best with ID-based pagination (numeric IDs that can be compared). For other pagination types:

Type	Adaptation
Cursor/token	Store cursor string instead of ID; can't compare numerically
Timestamp	Use last_timestamp column; compare as dates
Offset/limit	Store page number; resume from last saved page

See references/patterns.md for schemas and code examples.

Gotchas

• Save watermarks only after full success. If the process crashes mid-run, unsaved watermarks mean the next run re-fetches and deduplicates from scratch — no data loss, but potentially slow. Saving watermarks mid-run causes permanent gaps.
• Newest ID may not equal the highest numeric ID. Some APIs return IDs that are not monotonically increasing (e.g., snowflake IDs with clock drift). Always compare using the API's own ordering guarantees, not numeric comparison.
• Backfill mode must not overwrite the newest_id. A backfill run extends history backward; it should update only oldest_id. Overwriting newest_id during backfill causes duplicate fetches on the next forward update run.
• Rate-limit headers vary by API. Twitter uses x-rate-limit-reset; others use Retry-After. Check the specific API's response headers before implementing wait logic.