TopolearnResult lists need stable ordering and paging, and naive deep paging hits hard limits.
By default hits are ordered by _score. "sort": [{ "price": "asc" }, { "_score": "desc" }] sorts by fields; sort on keyword, numeric or date fields, never on analyzed text.
from/size paging ("from": 20, "size": 10 for page 3) is fine for the first pages. Each shard must produce from + size hits and the coordinating node merges them all, so cost grows with depth, and requests beyond index.max_result_window (10,000 by default) are rejected.
For deep pagination or exporting results use search_after:
- open a point in time (
POST /index/_pit?keep_alive=1m) so pages see a consistent snapshot even while documents change, - sort by your fields plus the implicit
_shard_doctie-breaker that PIT searches add, - pass each last hit's
sortvalues assearch_afterin the next request, - close the PIT when done.
search_after can't jump to page 500 directly, which matches how infinite scroll and exports work anyway. The older scroll API still exists but is no longer recommended for deep pagination; prefer PIT with search_after.
track_total_hits controls exact counting: false skips it entirely for speed, true counts exactly at a cost.
Example
POST /orders/_pit?keep_alive=1m
# { "id": "46ToAwMDaWR5BXV1..." }
GET /_search
{
"size": 100,
"query": { "term": { "status": "paid" } },
"pit": { "id": "46ToAwMDaWR5BXV1...", "keep_alive": "1m" },
"sort": [ { "placed_at": "desc" } ]
}
# each hit ends with "sort": [1760700000000, 4294967298]
GET /_search
{
"size": 100,
"query": { "term": { "status": "paid" } },
"pit": { "id": "46ToAwMDaWR5BXV1...", "keep_alive": "1m" },
"sort": [ { "placed_at": "desc" } ],
"search_after": [1760700000000, 4294967298]
}
DELETE /_pit
{ "id": "46ToAwMDaWR5BXV1..." }Remember
- Sort on keyword, numeric or date fields
- from/size for shallow pages; limited to 10,000 by default
- Deep paging: point in time + search_after
- Close PITs when you're done
Watch out for
- Raising max_result_window to millions to keep using from/size, which loads the cluster heavily.