Measured in production
Wake times
A sleeping engine starts again on the next search. This is how long that takes on our production servers, over the last 30 days.
Each engine gets two numbers. The median is a typical wake: half were faster. The p95, short for 95th percentile, is the slow end: 95 in 100 wakes were faster.
Updated November 15, 2026, 17:07 UTC.
Median wake with the data on the server
The large number is the median. Under it, the p95: 95 in 100 wakes were faster than that.
- Visor
- 150 msp95 293 ms · 70 wakes
- Typesense
- 65 msp95 164 ms · 27 wakes
- Meilisearch
- 168 msp95 376 ms · 111 wakes
- Failed wakes
- 0Across every engine
By engine
Median and 95th percentile (p95)
Each line runs from the median wake to the p95.
Meilisearch
- Data on the server111 wakesMedian 168 msp95 376 ms111
- Data from storage122 wakesMedian 811 msp95 2.0 s122
TypesenseResumes from a memory snapshot
- Data on the server27 wakesMedian 65 msp95 164 ms27
Visor
- Data on the server70 wakesMedian 150 msp95 293 ms70
- Data from storage21 wakesMedian 884 msp95 2.4 s21
ElasticsearchSleeps only on forked branches; on main it is always on and never waits
- Data on the server1 wakeMedian 11 sp95 too fewToo few1
- Data from storage6 wakesMedian 12 sp95 too fewToo few6
All the numbers
All the numbers
| Engine and wake path | Wakes | Median | p95 | Slowest | First answer, median / p95 |
|---|---|---|---|---|---|
| MeilisearchData on the server | 111 | 168 ms | 376 ms | 1.2 s | 183 ms / 415 ms |
| MeilisearchData from storage | 122 | 811 ms | 2.0 s | 5.1 s | 1.1 s / 3.3 s |
| TypesenseData on the server | 27 | 65 ms | 164 ms | 171 ms | 70 ms / 176 ms |
| VisorData on the server | 70 | 150 ms | 293 ms | 2.6 s | 208 ms / 366 ms |
| VisorData from storage | 21 | 884 ms | 2.4 s | 4.2 s | 1.2 s / 2.5 s |
| ElasticsearchData on the server | 1 | 11 s | Too few | 11 s | 11 s / Too few |
| ElasticsearchData from storage | 6 | 12 s | Too few | 15 s | 12 s / Too few |
MeilisearchData on the server111 wakes
- Median
- 168 ms
- p95
- 376 ms
- Slowest
- 1.2 s
- First answer
- 183 ms / 415 ms
MeilisearchData from storage122 wakes
- Median
- 811 ms
- p95
- 2.0 s
- Slowest
- 5.1 s
- First answer
- 1.1 s / 3.3 s
TypesenseData on the server27 wakes
- Median
- 65 ms
- p95
- 164 ms
- Slowest
- 171 ms
- First answer
- 70 ms / 176 ms
VisorData on the server70 wakes
- Median
- 150 ms
- p95
- 293 ms
- Slowest
- 2.6 s
- First answer
- 208 ms / 366 ms
VisorData from storage21 wakes
- Median
- 884 ms
- p95
- 2.4 s
- Slowest
- 4.2 s
- First answer
- 1.2 s / 2.5 s
ElasticsearchData on the server1 wake
- Median
- 11 s
- p95
- Too few
- Slowest
- 11 s
- First answer
- 11 s / Too few
ElasticsearchData from storage6 wakes
- Median
- 12 s
- p95
- Too few
- Slowest
- 15 s
- First answer
- 12 s / Too few
Failed wakes in the same 30 days: Meilisearch 0, Typesense 0, Visor 0, Elasticsearch 0.
What the numbers mean
- Wake time
- From the request that started the wake reaching our server to the engine being ready to answer. Your network is not included.
- First answer
- From the same moment to the first byte of the answer to that request. From storage it can come later, because the request may read data that is still being fetched.
- Data on the server
- The engine's disk is still on the server it slept on, so it only has to start. Sleeping Typesense resumes from a memory snapshot instead. This is almost every wake.
- Data from storage
- The server holds no copy of the engine's disk, after a move to another server or when a server needed the room. An engine on bucket storage starts while its data streams in. Any other engine is restored from its latest backup first.
- Median and p95
- Half of the wakes were faster than the median, and 95 in 100 were faster than the p95. A p95 needs at least 10 wakes; with fewer, the page says so.
What is counted
- Every wake of a sleeping engine that a request waited for.
- Left out: wakes nobody waited for, such as a warm call, and restarts after an engine crashed. Until November 30, 2026, the Meilisearch numbers still include four such restarts from our failover drills in late October, recorded as wakes before we could tell the two apart.
- Most wakes so far are of our own test and drill engines, which hold small indexes. A much larger index can take longer to start.
During a wake
The search waits, then gets its answer as usual. We hold it for up to 10 seconds on Free and 25 seconds on paid plans on an engine's own endpoint, and 30 seconds through the REST API on paid plans. The cold starts docs say what happens past that.
An engine set to Always on never sleeps, so it never wakes. See pricing for what that costs on each plan.