Migrate IndexFaces and SearchFacesByImage from AWS Rekognition

If you run face search on AWS Rekognition, the friction usually isn't the API — it's everything around it: raising per-operation TPS quotas by support ticket, no console to browse a collection, JPEG/PNG-only ingestion, and a per-image price that adds up. SightRadar is a dedicated face recognition API that mirrors Rekognition's face operations, so you can move without rewriting your integration.

What actually transfers (and what doesn't)

Your code transfers: the request and response shapes for the face-collection operations match. Your face vectors do not — embeddings are engine-specific, so you re-index the source images against the new engine. Plan the migration around re-indexing from your own image store, not exporting vectors from AWS.

Step 1 — Point your existing client at the new endpoint

SightRadar exposes a native REST API at https://api.sightradar.com with Authorization: Bearer <key>. Keys are created in the console and shown once. Here's the smallest possible index-then-search against the native API:

import os, requests

BASE = "https://api.sightradar.com"
H = {"Authorization": f"Bearer {os.environ['SR_API_KEY']}"}

# Create a collection (free — only face ops are billable)
requests.post(f"{BASE}/v1/collections", headers=H,
              json={"collection_id": "event-2026"})

# Index a photo. Pass your own photoId so you can map matches back later.
requests.post(f"{BASE}/v1/collections/event-2026/index", headers=H,
              json={"url": "https://cdn.example.com/img/001.jpg",
                    "photoId": "img-001"})

# Search with a selfie. Returns photo_ids ranked by similarity.
r = requests.post(f"{BASE}/v1/collections/event-2026/search", headers=H,
                  json={"url": "https://cdn.example.com/selfie.jpg", "limit": 10})
print(r.json())                              # ranked matches (JSON body)
print(r.headers["X-Credits-Remaining"])      # remaining credits (response header)

If you prefer to keep your AWS SDK call sites unchanged, use the open-source sightradar-rekognition-shim, which maps IndexFaces/SearchFacesByImage/CompareFaces onto the SightRadar endpoint so your boto3-style code keeps working. See the migration page for the side-by-side.

Step 2 — Map face ids back to your records

A production integration is mostly bookkeeping. Store the mapping between the API's face/photo identifiers and your own domain records so a search result becomes a person, image, and tenant:

create table face_index (
  tenant_id   text not null,
  photo_id    text not null,   -- your id, passed as photoId at index time
  source_url  text not null,
  face_count  int  not null,
  indexed_at  timestamptz not null default now(),
  primary key (tenant_id, photo_id)
);
-- one collection per tenant (or per event) keeps searches isolated

Step 3 — Zero-downtime backfill with dual writes

1. Dual-write new index calls to both Rekognition and SightRadar while you backfill. Reads still serve from Rekognition.
2. Backfill historical images by re-indexing from your object store into SightRadar collections. This is the bulk of the work; use batch indexing for the lower per-photo rate.
3. Validate by running a fixed set of known selfies against both engines and comparing the ranked results — not raw scores, which aren't comparable across engines.
4. Cut reads over to SightRadar once parity holds, keeping dual writes for a grace period.
5. Decommission the Rekognition path after the grace window.

Step 4 — Rollback plan

Because you kept dual writes and never deleted the Rekognition collection during the grace period, rollback is a read-path flip back to AWS. Keep the identifier mapping stable across both engines so no re-mapping is needed if you revert.

What this migration does not do

• It does not make your product legally compliant. You still capture consent, isolate tenants, and honor deletion — see Trust & responsible use.
• It does not provide liveness / anti-spoofing. Face matching alone does not stop a printed-photo or replay attack; add liveness if your use case needs it.
• It does not migrate vectors. You re-index source images; plan storage and time for the backfill.