Latency
How long it takes for something that happened to appear here, per path, measured continuously. The target is different for each path and the reason is printed beside it — a single figure across all of them was wrong, and wrong in the direction that flatters nobody: it was measured on the fastest path and then applied to the slowest, where it cannot be met by construction.
| Path and why it costs what it costs | Target | Median | p90 | Rows | Excluded | Verdict |
|---|---|---|---|---|---|---|
| Bitcoin whales A block arrives as a push ANNOUNCEMENT about a second after it is mined; the 1.8MB body must then be pulled from another host and walked for large transfers. That pull is inside this number and is never netted out of it — quoting the socket's own latency here would be reporting the announcement and calling it the event. Structural, not a defect. | ≤35s | 38.1s | 55.4s | 122 | 23 / 0 | over targetp90 is over it |
| Stories The pipeline path, unchanged: a source publishes, a tick picks it up, the story is written, the site is rebuilt and deployed. Measured from the source's own stamp to our first publication. | ≤10 min | 9.7 min | 170 min | 48 | 139 unmeasurable 126 were published before this wire recorded source timestamps (2026-08-12); 10 arrived on a feed that stated no time; 3 carried a source clock ahead of our own publish. The exclusion is an artefact of when the field was added, not a slow path being left out — but it does mean this median describes the wire as it runs now, and we have no measurement of the days before 2026-08-12. | within target 27 of 48 inside |
What “excluded” means
Two kinds of row are counted and then left out of the percentiles, and the counts are in the table above so you can see how many:
- Source clock ahead. A Bitcoin block carries the miner’s timestamp, and consensus only requires it to beat the median of the last eleven blocks and to stay under network time plus two hours — so miners routinely stamp ahead. Counting those gave a median of minus 158 seconds: we appeared to detect transfers two and a half minutes before they happened. That is a measurement of two clocks, not of this system.
- Opening backlog. A socket answers a subscription with its last several blocks. On a cold start that is history and on a reconnect it is the outage — one such row once reported 1,145 seconds for a path that took 24.
The verdict is on the median. Where the p90 is over the bar it says so, and that is the number worth arguing about: half of everything is faster than a median by definition, and the slow tail is what somebody actually waited for. The story row’s exclusions are posts carrying no source timestamp to measure from — on-chain stories, the Market Wrap, desk columns and anything published before the field existed — not slow ones dropped.
Daily median, last 30 days
1 of 30 days carry measurements. The live meter is a rolling window of the last 2,000 events — about two days — so the empty bars are days this record does not reach, not days nothing happened. They are drawn rather than dropped: a chart that hid them would imply a month of measurements we do not have.
When it was slower than this
- 2026-08-17 · mempool.space stopped answering this server, and Bitcoin went quiet · 63 min — the post-mortem
- 2026-08-17 · Ticks committed the news and deployed none of it · 33 min — the post-mortem
- 2026-08-13 · The wire stopped publishing for seven hours and nothing recovered it · 7h 16m — the post-mortem
Uptime and incidents · Accuracy · the on-chain ledger these are measured on