Dataset Viewer
Auto-converted to Parquet Duplicate
run_id
int64
run_date
timestamp[s]
call
string
n
int64
failures
int64
attempts
int64
p95_all_attempts_ms
float64
target_p95_ms
int64
met
bool
status
string
min_ms
float64
p50_ms
float64
p95_ms
float64
p99_ms
float64
max_ms
float64
cycles
int64
source_path
string
42,563,147,792
2026-11-09T00:00:00
order
60
0
60
1,227.49
1,000
false
MISSED
563.37
759.42
1,227.49
1,473.48
1,473.48
60
docs/evidence/latency/per-call-measure-sandbox.json
42,563,147,792
2026-11-09T00:00:00
authorize
60
0
60
1,538.31
1,000
false
MISSED
1,028.12
1,308.01
1,538.31
1,911.93
1,911.93
60
docs/evidence/latency/per-call-measure-sandbox.json
42,563,147,792
2026-11-09T00:00:00
capture
59
3
62
3,724.76
1,000
false
MISSED
1,614.21
2,093.26
2,718.84
3,724.76
3,724.76
60
docs/evidence/latency/per-call-measure-sandbox.json
42,563,147,792
2026-11-09T00:00:00
void
60
0
60
1,249.84
1,000
false
MISSED
702.19
947.66
1,249.84
1,458.67
1,458.67
60
docs/evidence/latency/per-call-measure-sandbox.json

gavel-latency

Honest labels: sandbox. Measured against the PayPal sandbox, not production, and with few samples. Sandbox latency is not production latency. Latest run (42563147792): hammer to captured p95 3766.67 ms against 2500 ms: MISSED (49 of 50 completed); hammer to settled p95 6927.43 ms against 6000 ms: MISSED (49 of 50 completed). Read status and met per row; a row with failures is never MET.

What it is

Latency of the payment calls an auction settlement makes, measured against the real PayPal sandbox: the time from the hammer to the winner being captured, from the hammer to the whole settlement being finished (winner captured, every other hold released), and the per-call latency of order, authorize, capture and void.

The tables hold summary statistics (min, p50, p95, p99, max in milliseconds, attempt and failure counts). The raw per-call samples were not kept by the measuring job. Percentiles are over the attempts that completed; p95_all_attempts_ms counts every failed or timed-out attempt as the worst case.

How it was made

  • Settlement spans: run https://github.com/harshbhx/gavel-auction/actions/runs/42346811713 (20 settlements, 20 timed, source docs/evidence/measure/measure-settlement.json) and run https://github.com/harshbhx/gavel-auction/actions/runs/42563147792 (50 settlements, 49 timed, source docs/evidence/latency/settlement-measure-settlement.json). Each settlement has 6 holds, one winner, and voids five holds three at a time. A sandbox buyer approves every order in a browser; the approval is not timed.
  • API calls: run https://github.com/harshbhx/gavel-auction/actions/runs/42563147792, 60 sequential cycles (source docs/evidence/latency/per-call-measure-sandbox.json). The same numbers are in docs/GATES.md of the repository.
  • The measuring job uploads numbers only (counts, milliseconds, status counts). The build script reads the redacted summaries committed under docs/evidence directly and only reshapes them; there are no hand-copied inputs. The targets are the ones in docs/GATES.md and are never changed to fit a result.

Provenance

Built from commit 65be30bfccc88c8fb836d48f11432423460933f5 of https://github.com/harshbhx/gavel-auction by https://github.com/harshbhx/gavel-auction/actions/runs/43196464384. The same values are in MANIFEST.json.

MANIFEST.json lists the sha256 and size of every file and the row count of every table. python verify.py recomputes the hashes and fails on any difference.

Fields

settlement_spans.jsonl / .csv

One row per run and span (4 rows).

Field Type Meaning
run_id integer GitHub Actions run id
run_label string Settlements attempted in the run, e.g. 50 settlements
run_date string or null Date of the run (UTC), from docs/GATES.md
span string hammer_to_captured or hammer_to_settled
n integer Settlements that completed and are in the percentiles
failures integer Settlements that failed or timed out (not in the percentiles)
attempts integer n plus failures
p95_all_attempts_ms float or null p95 over all attempts with every failure counted as the worst case; an upper bound read from the summary, null when a failed attempt sits at the 95th percentile (unbounded)
target_p95_ms integer Target for this span from docs/GATES.md
met, status boolean, string MET, MISSED or UNPROVEN (target met on the completed attempts but some attempts failed, so not met)
min_ms, p50_ms, p95_ms, p99_ms, max_ms float Latency statistics in milliseconds over the completed attempts
holds_per_settlement integer Holds per settlement
void_concurrency integer Voids in flight at a time
mismatches, errors integer Ledger mismatches and errors reported by the run
source_path string Committed evidence file the row was reshaped from

api_calls.jsonl / .csv

One row per run and PayPal call type (4 rows).

Field Type Meaning
run_id, run_date integer, string or null Run and date
call string order, authorize, capture or void
n integer Calls that completed and are in the percentiles
failures integer Calls that failed
attempts, cycles integer n plus failures; sequential cycles in the run
p95_all_attempts_ms float or null p95 with every failed call counted as the worst case (bound; null = unbounded)
target_p95_ms, met, status integer, boolean, string Per-call p95 budget and whether it was met (MET, MISSED, UNPROVEN)
min_ms, p50_ms, p95_ms, p99_ms, max_ms float Latency statistics in milliseconds over the completed calls
source_path string Committed evidence file the row was reshaped from

Intended use

Sizing timeouts and retry budgets for payment flows, sanity-checking latency targets, and as a reference point for how slow a sandbox payment provider can be.

Limitations

  • Sandbox only, from a GitHub-hosted runner, at one time of day. Production will differ.
  • Small n: 20 settlements and 50 settlements, 60 call cycles. Treat percentiles as indicative, not as service-level numbers.
  • Failed or timed-out attempts are not in the percentiles (run 42563147792 hammer_to_captured: 1 of 50 attempts failed or timed out; run 42563147792 hammer_to_settled: 1 of 50 attempts failed or timed out; run 42563147792 capture: 3 of 62 attempts failed or timed out). They are usually the slowest attempts, so the reported p95 can understate the tail; use p95_all_attempts_ms and status.
  • Only runs whose redacted summaries are committed under docs/evidence are included. A 3-settlement pilot (run 42346020463) and a 10-cycle per-call run (42047543821) were never committed as evidence files; docs/GATES.md still records them.
  • Summary statistics only; no per-sample data.
  • The sandbox rate limits shaped how the runs were paced.

License

CC-BY-4.0 (Creative Commons Attribution 4.0 International). You may copy, adapt and redistribute the data, including commercially, if you credit the project (see Citation). The licence covers the data files in this dataset only; it grants no right to the marks or services of PayPal, Deepgram, Groq or any other third party named in it.

Simulated or real

Labels: sandbox. Read the label note at the top of this card before using any number. Nothing here is production data, and no real person or real money is involved.

Citation

@dataset{gavel_gavel_latency_2026,
  title  = {gavel-latency},
  author = {Bhanushali, Harsh},
  year   = {2026},
  url    = {https://huggingface.co/datasets/harshbhx/gavel-latency},
  note   = {Generated from https://github.com/harshbhx/gavel-auction}
}

Links

Contact

Open an issue at https://github.com/harshbhx/gavel-auction/issues. The code that builds this dataset is in the scripts/hf folder of that repository.

Downloads last month
138

Collection including harshbhx/gavel-latency