Scatter, reduce, fetch — the two-phase query-then-fetch search

Every _search is two round trips — phase 1 scatters to one copy of every shard for top-(from+size) (doc_id, score) tuples and phase 2 fetches bodies for just the winners — which is why from + size is bounded at 10 000.

Scene 07

Scatter, reduce, fetch

Watch
Try it
Predict
Capture

What to watch for

The receiving node becomes the conductor. Watch phase 1 fan out 'distributed systems' to one copy of every shard, the coordinator merge the local top-Ks into a global top-N, then phase 2 fetch _source ONLY from the shards that actually contributed a winner.

Implementation

Highlighted lines are the ones running in the diagram right now.

Coordinator.search

phase 1 scatter for top-(from+size); phase 2 fetch winners

1def search(query, from, size):
2    # phase 1 — scatter (doc_id, score) only
3    results = []
4    for shard in all_primary_shards:
5        copy = adaptive_replica_selection(shard)
6        results.append(copy.localSearch(query, from + size))
7    top = merge(results, from, size)
8    # phase 2 — fetch _source for global winners only
9    bodies = [shard_of(h).fetch(h.doc_id) for h in top]
10    return assemble(top, bodies)

Shard.localSearch

BM25-rank local matches; ship (doc_id, score) only

1def localSearch(query, top_k):
2    pq = PriorityQueue(top_k)
3    for doc_id in postings_for(query):
4        score = bm25(query, doc_id)
5        pq.offer((doc_id, score))
6    return pq.drain()   # no _source yet

Coordinator.merge

buffer N × (from+size) candidates, then slice

1def merge(per_shard, from, size):
2    # buffer = num_shards * (from + size)
3    # this is the from + size <= 10 000 wall
4    pool = []
5    for hits in per_shard:
6        pool.extend(hits)
7    pool.sort(key=score, reverse=True)
8    return pool[from : from + size]

Where this sits in Build a distributed search engine (Elasticsearch / OpenSearch style)

Scene 07 of 12. _search is two round trips: scatter top-(from+size) doc-ids+scores from every shard, reduce, then fetch only the global winners — which is where the 10,000 page wall comes from.

All 12 scenes in Build a distributed search engine (Elasticsearch / OpenSearch style) · Every curriculum

Designs that use this

Built with Arqly

Every scene in Build a distributed search engine (Elasticsearch / OpenSearch style) builds on the one before it.All 12 Build a distributed search engine (Elasticsearch / OpenSearch style) scenes