gitlore/docs/api-efficiency-findings.md

---
plan: true
title: "api-efficiency-findings"
status: drafting
iteration: 0
target_iterations: 8
beads_revision: 0
related_plans: []
created: 2026-02-07
updated: 2026-02-07
---

# API Efficiency & Observability Findings

> **Status:** Draft - working through items
> **Context:** Audit of gitlore's GitLab API usage, data processing, and observability gaps
> **Interactive reference:** `api-review.html` (root of repo, open in browser)

---

## Checkpoint 3 Alignment

Checkpoint 3 (`docs/prd/checkpoint-3.md`) introduces `lore sync` orchestration, document generation, and search. Several findings here overlap with that work. This section maps the relationship so effort isn't duplicated and so CP3 implementation can absorb the right instrumentation as it's built.

### Direct overlaps (CP3 partially addresses)

| Finding | CP3 coverage | Remaining gap |
|---------|-------------|---------------|
| **P0-1** sync_runs never written | `lore sync` step 7 says "record sync_run". `SyncResult` struct defined with counts. | Only covers the new `lore sync` command. Existing `lore ingest` still won't write sync_runs. Either instrument `lore ingest` separately or have `lore sync` subsume it entirely. |
| **P0-2** No timing | `print_sync` captures wall-clock `elapsed_secs` / `elapsed_ms` in robot mode JSON `meta` envelope. | Wall-clock only. No per-phase, per-API-call, or per-DB-write breakdown. The `SyncResult` struct has counts but no duration fields. |
| **P2-1** Discussion full-refresh | CP3 introduces `pending_discussion_fetches` queue with exponential backoff and bounded processing per sync. Structures the work better. | Same full-refresh strategy per entity. The queue adds retry resilience but doesn't reduce the number of API calls for unchanged discussions. |

### Different scope (complementary, no overlap)

| Finding | Why no overlap |
|---------|---------------|
| **P0-3** metrics_json schema | CP3 doesn't reference the `metrics_json` column. `SyncResult` is printed/returned but not persisted there. |
| **P0-4** Discussion sync telemetry columns | CP3's queue system (`pending_discussion_fetches`) is a replacement architecture. The existing per-MR telemetry columns (`discussions_sync_attempts`, `_last_error`) aren't referenced in CP3. Decide: use CP3's queue table or wire up the existing columns? |
| **P0-5** Progress events lack timing | CP3 lists "Progress visible during long syncs" as acceptance criteria but doesn't spec timing in events. |
| **P1-\*** Free data capture | CP3 doesn't touch GitLab API response field coverage at all. These are independent. |
| **P2-2** Keyset pagination (GitLab API) | CP3 uses keyset pagination for local SQLite queries (document seeding, embedding pipelines). Completely different from using GitLab API keyset pagination. |
| **P2-3** ETags | Not mentioned in CP3. |
| **P2-4** Labels enrichment | Not mentioned in CP3. |
| **P3-\*** Structural improvements | Not in CP3 scope. |

### Recommendation

CP3's `lore sync` orchestrator is the natural integration point for P0 instrumentation. Rather than retrofitting `lore ingest` separately, the most efficient path is:

1. Build P0 timing instrumentation as a reusable layer (e.g., a `SyncMetrics` struct that accumulates phase timings)
2. Wire it into the CP3 `run_sync` implementation as it's built
3. Have `run_sync` persist the full metrics (counts + timing) to `sync_runs.metrics_json`
4. Decide whether `lore ingest` becomes a thin wrapper around `lore sync --no-docs --no-embed` or stays separate with its own sync_runs recording

This avoids building instrumentation twice and ensures the new sync pipeline is observable from day one.

### Decision: `lore ingest` goes away

`lore sync` becomes the single command for all data fetching. First run does a full fetch (equivalent to today's `lore ingest`), subsequent runs are incremental via cursors. `lore ingest` becomes a hidden deprecated alias.

Implications:
- P0 instrumentation only needs to be built in one place (`run_sync`)
- CP3 Gate C owns the sync_runs lifecycle end-to-end
- The existing `lore ingest issues` / `lore ingest mrs` code becomes internal functions called by `run_sync`, not standalone CLI commands
- `lore sync` always syncs everything: issues, MRs, discussions, documents, embeddings (with `--no-embed` / `--no-docs` to opt out of later stages)

---

## Implementation Sequence

### Phase A: Before CP3 (independent, enriches data model)

**Do first.** Migration + struct changes only. No architectural dependency. Gets richer source data into the DB before CP3's document generation pipeline locks in its schema.

1. **P1 batch: free data capture** - All ~11 fields in a single migration. `user_notes_count`, `upvotes`, `downvotes`, `confidential`, `has_conflicts`, `blocking_discussions_resolved`, `merge_commit_sha`, `discussion_locked`, `task_completion_status`, `issue_type`, `issue references`.
2. **P1-10: MR milestones** - Reuse existing issue milestone transformer. Slightly more work, same migration.

### Phase B: During CP3 Gate C (`lore sync`)

**Build instrumentation into the sync orchestrator as it's constructed.** Not a separate effort.

3. **P0-1 + P0-2 + P0-3** - `SyncMetrics` struct accumulating phase timings. `run_sync` writes to `sync_runs` with full `metrics_json` on completion.
4. **P0-4** - Decide: use CP3's `pending_discussion_fetches` queue or existing per-MR telemetry columns. Wire up the winner.
5. **P0-5** - Add `elapsed_ms` to `*Complete` progress event variants.
6. **Deprecate `lore ingest`** - Hidden alias pointing to `lore sync`. Remove from help output.

### Phase C: After CP3 ships, informed by real metrics

**Only pursue items that P0 data proves matter.**

7. **P2-1: Discussion optimization** - Check metrics_json from real runs. If discussion phase is <10% of wall-clock, skip.
8. **P2-2: Keyset pagination** - Check primary fetch timing on largest project. If fast, skip.
9. **P2-4: Labels enrichment** - If label colors are needed for any UI surface.

### Phase D: Future (needs a forcing function)

10. **P3-1: Users table** - When a UI needs display names / avatars.
11. **P2-3: ETags** - Only if P2-1 doesn't sufficiently reduce discussion overhead.
12. **P3-2/3/4: GraphQL, Events API, Webhooks** - Architectural shifts. Only if pull-based sync hits a scaling wall.

---

## Priority 0: Observability (prerequisite for everything else)

We can't evaluate any efficiency question without measurement. Gitlore has no runtime performance instrumentation. The infrastructure for it was scaffolded (sync_runs table, metrics_json column, discussion sync telemetry columns) but never wired up.

### P0-1: sync_runs table is never written to

**Location:** Schema in `migrations/001_initial.sql:25-34`, read in `src/cli/commands/sync_status.rs:69-72`

The table exists and `lore status` reads from it, but no code ever INSERTs or UPDATEs rows. The entire audit trail is empty.

```sql
-- Exists in schema, never populated
CREATE TABLE sync_runs (
  id INTEGER PRIMARY KEY,
  started_at INTEGER NOT NULL,
  heartbeat_at INTEGER NOT NULL,
  finished_at INTEGER,
  status TEXT NOT NULL,        -- 'running' | 'succeeded' | 'failed'
  command TEXT NOT NULL,
  error TEXT,
  metrics_json TEXT            -- never written
);
```

**What to do:** Instrument the ingest orchestrator to record sync runs. Each `lore ingest issues` / `lore ingest mrs` invocation should:
- INSERT a row with status='running' at start
- UPDATE with status='succeeded'/'failed' + finished_at on completion
- Populate metrics_json with the IngestProjectResult / IngestMrProjectResult counters

### P0-2: No operation timing anywhere

**Location:** Rate limiter in `src/gitlab/client.rs:20-65`, orchestrator in `src/ingestion/orchestrator.rs`

`Instant::now()` is used only for rate limiter enforcement. No operation durations are measured or logged. We don't know:

- How long a full issue ingest takes
- How long discussion sync takes per entity
- How long individual API requests take (network latency)
- How long database writes take per batch
- How long rate limiter sleeps accumulate to
- How long pagination takes across pages

**What to do:** Add timing instrumentation at these levels:

| Level | What to time | Where |
|-------|-------------|-------|
| **Run** | Total ingest wall-clock time | orchestrator entry/exit |
| **Phase** | Primary fetch vs discussion sync | orchestrator phase boundaries |
| **API call** | Individual HTTP request round-trip | client.rs request method |
| **DB write** | Transaction duration per batch | ingestion store functions |
| **Rate limiter** | Cumulative sleep time per run | client.rs acquire() |

Store phase-level and run-level timing in `metrics_json`. Log API-call-level timing at debug level.

### P0-3: metrics_json has no defined schema

**What to do:** Define what goes in there. Strawman based on existing IngestProjectResult fields plus timing:

```json
{
  "wall_clock_ms": 14200,
  "phases": {
    "primary_fetch": {
      "duration_ms": 8400,
      "api_calls": 12,
      "items_fetched": 1143,
      "items_upserted": 87,
      "pages": 12,
      "rate_limit_sleep_ms": 1200
    },
    "discussion_sync": {
      "duration_ms": 5800,
      "entities_checked": 87,
      "entities_synced": 14,
      "entities_skipped": 73,
      "api_calls": 22,
      "discussions_fetched": 156,
      "notes_upserted": 412,
      "rate_limit_sleep_ms": 2200
    }
  },
  "db": {
    "labels_created": 3,
    "raw_payloads_stored": 87,
    "raw_payloads_deduped": 42
  }
}
```

### P0-4: Discussion sync telemetry columns are dead code

**Location:** `merge_requests` table columns: `discussions_sync_last_attempt_at`, `discussions_sync_attempts`, `discussions_sync_last_error`

These exist in the schema but are never read or written. They were designed for tracking retry behavior on failed discussion syncs.

**What to do:** Wire these up during discussion sync. On attempt: set last_attempt_at and increment attempts. On failure: set last_error. On success: reset attempts to 0. This provides per-entity visibility into discussion sync health.

### P0-5: Progress events carry no timing

**Location:** `src/ingestion/orchestrator.rs:28-53`

ProgressEvent variants (`IssueFetched`, `DiscussionSynced`, etc.) carry only counts. Adding elapsed_ms to at least `*Complete` variants would give callers (CLI progress bars, robot mode output) real throughput numbers.

---

## Priority 1: Free data capture (zero API cost)

These fields are already in the API responses gitlore receives. Storing them requires only Rust struct additions and DB column migrations. No additional API calls.

### P1-1: user_notes_count (Issues + MRs)

**API field:** `user_notes_count` (integer)
**Value:** Could short-circuit discussion re-sync. If count hasn't changed, discussions probably haven't changed either. Also useful for "most discussed" queries.
**Effort:** Add field to serde struct, add DB column, store during transform.

### P1-2: upvotes / downvotes (Issues + MRs)

**API field:** `upvotes`, `downvotes` (integers)
**Value:** Engagement metrics for triage. "Most upvoted open issues" is a common query.
**Effort:** Same pattern as above.

### P1-3: confidential (Issues)

**API field:** `confidential` (boolean)
**Value:** Security-sensitive filtering. Important to know when exposing issue data.
**Effort:** Low.

### P1-4: has_conflicts (MRs)

**API field:** `has_conflicts` (boolean)
**Value:** Identify MRs needing rebase. Useful for "stale MR" detection.
**Effort:** Low.

### P1-5: blocking_discussions_resolved (MRs)

**API field:** `blocking_discussions_resolved` (boolean)
**Value:** MR readiness indicator without joining the discussions table.
**Effort:** Low.

### P1-6: merge_commit_sha (MRs)

**API field:** `merge_commit_sha` (string, nullable)
**Value:** Trace merged MRs to specific commits in git history.
**Effort:** Low.

### P1-7: discussion_locked (Issues + MRs)

**API field:** `discussion_locked` (boolean)
**Value:** Know if new comments can be added. Useful for robot mode consumers.
**Effort:** Low.

### P1-8: task_completion_status (Issues + MRs)

**API field:** `task_completion_status` (object: `{count, completed_count}`)
**Value:** Track task-list checkbox progress without parsing markdown.
**Effort:** Low. Store as two integer columns or a small JSON blob.

### P1-9: issue_type (Issues)

**API field:** `issue_type` (string: "issue" | "incident" | "test_case")
**Value:** Distinguish issues vs incidents vs test cases for filtering.
**Effort:** Low.

### P1-10: MR milestone (MRs)

**API field:** `milestone` (object, same structure as on issues)
**Current state:** Milestones are fully stored for issues but completely ignored for MRs.
**Value:** "Which MRs are in milestone X?" Currently impossible to query locally.
**Effort:** Medium - reuse existing milestone transformer from issue pipeline.

### P1-11: Issue references (Issues)

**API field:** `references` (object: `{short, relative, full}`)
**Current state:** Stored for MRs (`references_short`, `references_full`), dropped for issues.
**Value:** Cross-project issue references (e.g., `group/project#42`).
**Effort:** Low.

---

## Priority 2: Efficiency improvements (requires measurement from P0 first)

These are potential optimizations. **Do not implement until P0 instrumentation proves they matter.**

### P2-1: Discussion full-refresh strategy

**Current behavior:** When an issue/MR's `updated_at` advances, ALL its discussions are deleted and re-fetched from scratch.

**Potential optimization:** Use `user_notes_count` (P1-1) to detect whether discussions actually changed. Skip re-sync if count is unchanged.

**Why we need P0 first:** The full-refresh may be fast enough. Since we already fetch the data from GitLab, the DELETE+INSERT is just local SQLite I/O. If discussion sync for a typical entity takes <100ms locally, this isn't worth optimizing. We need the per-entity timing from P0-2 to know.

**Trade-offs to consider:**
- Full-refresh catches edited and deleted notes. Incremental would miss those.
- `user_notes_count` doesn't change when notes are edited, only when added/removed.
- Full-refresh is simpler to reason about for consistency.

### P2-2: Keyset pagination

**Current behavior:** Offset-based (`page=N&per_page=100`).
**Alternative:** Keyset pagination (`pagination=keyset`), O(1) per page instead of O(N).

**Why we need P0 first:** Only matters for large projects (>10K issues). Most projects will never hit enough pages for this to be measurable. P0 timing of pagination will show if this is a bottleneck.

**Note:** Gitlore already parses `Link` headers for next-page detection, which is the client-side mechanism keyset pagination uses. So partial support exists.

### P2-3: ETag / conditional requests

**Current behavior:** All requests are unconditional.
**Alternative:** Cache ETags, send `If-None-Match`, get 304s back.

**Why we need P0 first:** The cursor-based sync already avoids re-fetching unchanged data for primary resources. ETags would mainly help with discussion re-fetches where nothing changed. If P2-1 (user_notes_count skip) is implemented, ETags become less valuable.

### P2-4: Labels API enrichment

**Current behavior:** Labels extracted from the `labels[]` string array in issue/MR responses. The `labels` table has `color` and `description` columns that may not be populated.
**Alternative:** Single call to `GET /projects/:id/labels` per project per sync to populate label metadata.
**Cost:** 1 API call per project per sync run.
**Value:** Label colors for UI rendering, descriptions for tooltips.

---

## Priority 3: Structural improvements (future consideration)

### P3-1: Users table

**Current state:** Only `username` stored. Author `name`, `avatar_url`, `web_url`, `state` are in every API response but discarded.
**Proposal:** Create a `users` table, upsert on every encounter. Zero API cost.
**Value:** Richer user display, detect blocked/deactivated users.

### P3-2: GraphQL API for field-precise fetching

**Current state:** REST API returns ~40-50 fields per entity. Gitlore uses ~15-23.
**Alternative:** GraphQL API allows requesting exactly the fields needed.
**Trade-offs:** Different pagination model, potentially less stable API, more complex client code. The bandwidth savings are real but likely minor compared to discussion re-fetch overhead.

### P3-3: Events API for lightweight change detection

**Endpoint:** `GET /projects/:id/events`
**Value:** Lightweight "has anything changed?" check before running full issue/MR sync. Could replace or supplement the cursor-based approach for very active projects.

### P3-4: Webhook-based push sync

**Endpoint:** `POST /projects/:id/hooks` (setup), then receive pushes.
**Value:** Near-real-time sync without polling cost. Eliminates all rate-limit concerns.
**Barrier:** Requires a listener endpoint, which changes the architecture from pull-only CLI to something with a daemon/server component.

---

## Working notes

_Space for recording decisions as we work through items._

### Decisions made

| Item | Decision | Rationale |
|------|----------|-----------|
| `lore ingest` | Remove. `lore sync` is the single entry point. | No reason to separate initial load from incremental updates. First run = full fetch, subsequent = cursor-based delta. |
| CP3 alignment | Build P0 instrumentation into CP3 Gate C, not separately. | Avoids building in two places. `lore sync` owns the full lifecycle. |
| P2 timing | Defer all efficiency optimizations until P0 metrics from real runs are available. | Can't evaluate trade-offs without measurement. |

### Open questions

- What's the typical project size (issue/MR count) for gitlore users? This determines whether keyset pagination (P2-2) matters.
- Is there a plan for a web UI or TUI? That would increase the value of P3-1 (users table) and P2-4 (label colors).