tapbench

tapbench is a small workload benchmark client for IVOA TAP asynchronous jobs. It schedules ADQL queries as a one-shot burst or at an offered rate, follows each admitted UWS job to a terminal phase, optionally streams its primary result, and reports latency and throughput. It is intentionally not a general-purpose load-testing framework.

Python 3.12 or newer is required.

Install

Install the command from PyPI with either uv or pipx:

uv tool install tapbench
# or
pipx install tapbench

From a checkout, uv can create the development environment and run the command:

uv sync
uv run tapbench --help

The TAP base URL normally comes from the nonempty TAPBENCH_URL environment variable. The service credentials must already be present as TAPBENCH_USER and TAPBENCH_PASSWORD; both must be nonempty. The command does not read dotenv files and does not accept credentials as command-line options.

Run a benchmark

export TAPBENCH_URL=https://example.org/tap
tapbench run examples/mppdb.large_data.sql

From a checkout, prefix the command with uv run.

With no scheduling options, every positional query is submitted once and all submissions are offered together at t=0. This makes the common shell-glob case a concurrent one-shot workload:

uv run --env-file .env.mppdb -- tapbench run mppdb.*.sql

The shell expands the glob before tapbench runs. Its expansion order becomes the query order, and the behavior applies equally to any explicit list of query files.

Each positional file must contain exactly one nonempty ADQL query. Multiple files may be supplied; opportunities rotate through them in argument order, including opportunities dropped locally. The queries themselves are never printed or stored in JSON; the report identifies each by its zero-based index, file name/path, and SHA-256 digest.

The run options are:

Run uv run tapbench run --help for the installed command's defaults.

Stage packaged examples

Installed wheels contain example families for the bq and mppdb deployments. List them or copy one family into a working directory without TAP credentials or network access:

tapbench examples list
tapbench examples stage bq queries
tapbench run queries/bq.large_data.sql

The destination defaults to the current directory and is created when needed. Staging preserves the packaged names and preflights every destination; if any would be overwritten, nothing is copied. Use --force to atomically replace existing files.

Scheduling behavior

There are three scheduling modes:

• With neither --rate nor --duration, tapbench offers every positional query exactly once at t=0 (burst_once). The query count must not exceed --max-inflight; an oversized burst is rejected before credentials or network activity instead of being reported as local drops.
• With --rate but no --duration, tapbench offers every positional query exactly once at absolute deadlines k / rate (paced_once).
• With --duration, tapbench retains fixed-duration scheduling (fixed_duration). The rate is the explicit --rate, or 1 job/s when it is omitted. Queries rotate for every opportunity until the duration ends.

Paced and fixed-duration operation is open-loop. Opportunity k has an absolute deadline of k / rate seconds from the monotonic scheduling epoch, beginning with opportunity zero at t=0. In fixed-duration mode, all deadlines strictly before duration are offered. Thus --rate 10 --duration 10 creates exactly 100 opportunities even when earlier queries are slow. Polls and result requests do not count toward the offered rate, and a late event-loop wake-up does not shift later deadlines. With multiple query files, opportunity k uses query index k % number_of_queries; a local drop still advances this rotation.

An opportunity is admitted only if an in-flight slot is immediately available. Otherwise it is recorded as local_scheduling_drop; it is never queued to run later. A --max-inflight slot is held from submission start through job creation, phase handling, optional result download, and attempted cleanup. Consequently, slow downloads or cleanup can legitimately cause later drops. This bounds local tasks and network work as well as remote nonterminal jobs.

After the scheduling interval, tapbench drains the bounded set of admitted operations. The job timeout covers waiting for a terminal phase, beginning after successful job creation. Result download and cleanup occur outside that job deadline and are instead bounded by per-operation request timeouts.

Job creation, job start, phase polling, and remote cleanup treat an explicit HTTP 429 response as transient server backpressure through one shared retry policy. tapbench honors a valid Retry-After delay (seconds or HTTP date), adds small jitter to avoid synchronized retries, and otherwise uses bounded jittered exponential backoff. Creation retries remain within the request timeout, start and polling retries remain within the original job deadline, and DELETE retries remain within a separate cleanup deadline. Ambiguous POST timeouts and transport failures are not retried. Each record and the terminal summary report stage-specific rate-limit counts and cumulative wait times.

By default, phase observation uses bounded UWS 1.1 blocking requests on the full job resource, such as GET {job}?WAIT=25&PHASE=EXECUTING. tapbench first observes the ordinary job document, then sends PHASE=RUN exactly once when that observation is PENDING; job creation never includes an inline PHASE=RUN. Each WAIT is shortened when necessary to fit both the remaining job deadline and the request timeout, including a response grace margin. WAIT responses are parsed incrementally with a 1 MiB total limit; query text, parameters, result URLs, owners, and raw XML are never retained.

WAIT support is learned from observed behavior per actual job origin. Ignored, rejected, malformed, repeatedly immediate/unchanged, or repeatedly timed-out WAIT requests fall back safely to jittered /phase polling. Three distinct-job negative samples temporarily suppress WAIT for an origin; suppression later grants one generation-safe probe so transient server behavior cannot disable WAIT permanently. Retry-After applies to both WAIT and fallback polling, and the absolute job deadline remains authoritative. HELD, SUSPENDED, and UNKNOWN use fallback polling. ARCHIVED is terminal and is reported as uws_archived without requesting a result.

When WAIT reports COMPLETED, tapbench closes that response and begins result discovery immediately in the same job coroutine, with no poll sleep, capability probe, or timing request in between. It does not opportunistically extract a result URL from the job document: download mode continues to request the standard {job}/results/result resource. Timing fields already delivered with the terminal phase are reused. If timing is incomplete, the best-effort job document retrieval runs only after the complete result download and before cleanup; in no-download mode it may run immediately after terminal observation. The same download-before-timing ordering applies when fallback polling observes completion.

UWS creationTime, startTime, and endTime yield query queue duration (startTime - creationTime) and execution duration (endTime - startTime). Missing, malformed, out-of-order, or unavailable timing remains null and does not change the outcome. A service's timestamp precision limits these measurements; a whole-second service may report zero for a subsecond query. The queue duration includes time spent PENDING before RUN, so it is not necessarily pure database scheduler wait.

Adaptive WAIT changes request load into connection load: an effective service receives fewer phase requests, while more HTTP connections can be held concurrently. Connections remain bounded by the run connector limit (about twice --max-inflight), and every job has at most one acquired or queued HTTP request at a time. Reports expose WAIT and fallback counts so runs remain interpretable.

Results and success

By default, a job is successful as soon as creation succeeds and COMPLETED is observed. No result-list, result-resource, redirect-target, or size-discovery request is made, and all result timing and byte fields are null.

With --download-results, a COMPLETED job is successful only after {job}/results/result has been completely consumed and no HTTP, TAP, or VOTable result error was detected. Response-header, first-byte, complete-body, and submission-to-download timings are recorded along with the byte count. The body is streamed rather than retained in memory. Cleanup failure is reported separately and does not change an otherwise successful operation.

Supplying a directory, as in --download-results results/, retains the first successful job for each query until the benchmark has fully drained. After the benchmark finish time is captured and progress reporting stops, tapbench fetches each representative result again and saves it as <query-file-stem>.vot; for example, bq.large_data.sql becomes bq.large_data.vot. Query files must have distinct base names when saving. These extra transfers do not alter the benchmark's per-job measurements, aggregate byte counts, or elapsed time. The directory must already exist, and tapbench refuses to overwrite an existing representative file. Each retained job is then cleaned up normally unless --keep-jobs was selected.

Result inspection is deliberately bounded and does not deserialize a complete VOTable. tapbench retains only a small prefix while streaming and detects a QUERY_STATUS=ERROR marker when it appears there. An error marker appearing only beyond that inspection bound cannot be detected.

Latency summaries ignore missing observations and use nearest-rank p50, p90, p95, and p99 values: sort the samples and select observation ceil(percentile * count). An unavailable metric is shown as unavailable and serialized as JSON null.

Throughput labels have distinct denominators:

• achieved submission rate is submission starts divided by the configured scheduling duration, or by elapsed scheduler-active time after interruption;
• job-creation throughput is created jobs divided by the time from the scheduling epoch through the last successful creation response;
• server execution throughput is completed jobs with valid UWS timing divided by the server work window from the earliest startTime to the latest endTime; it is independent of client polling and unavailable when the server timestamps produce a zero-width window;
• result-completion throughput, when downloading, is successful downloads divided by the download window from the first download start through the last completion;
• download throughput is successfully downloaded bytes divided by the span of that same download window. Download rates are unavailable when there are no successful downloads or the span is zero.

No throughput denominator includes phase-poll sleeps or post-completion cleanup. The report shows how many completed jobs supplied valid UWS timing because server execution throughput may use fewer samples than the completed count.

Progress and terminal output

Progress is operational telemetry; final metrics are recomputed from all records. It is written to stderr so the final human summary remains on stdout. On an interactive terminal, a Rich progress display refreshes in place:

⠹ Scheduling ━━━━━━━━━━━━━━━╺━━━━━━━━━━━━━━━━━━━━━━  40%  12.0/30.0s
  Completed  ━━━╺━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━  26/300 queries

  workload   offered 121  started 118  dropped 3   active 87  peak 91
  jobs       created 115  completed 26 error 2     aborted 0
  outcome    success 24   failed 5     downloaded 812 MiB
  UWS        polls 412    waits 97     changes 81  fallback polls 4  fallback jobs 2  WAIT disabled 0
  limits     submit 4     start 2      polling 11  cleanup 3

Once scheduling ends, the determinate bar becomes a pulsing drain indicator:

⠸ Draining   ━━━━━━━━━━━━━━━━━━━╺━━━━━━━━━━━━━━━━━━  87 active  42.4s elapsed
  Completed  ━━━━━━━━━━━━━━━━━━━━━━━━━━━╺━━━━━━━━━━  205/300 queries

  workload   offered 300  started 292  dropped 8   active 87  peak 200
  jobs       created 292  completed 205 error 0     aborted 0
  outcome    success 205  failed 8     downloaded 3.1 GiB
  UWS        polls 994    waits 246    changes 246 fallback polls 0  fallback jobs 0  WAIT disabled 0
  limits     submit 0     start 0      polling 0   cleanup 0

The green completion bar advances only for jobs reaching UWS COMPLETED; its total is the run's exact planned opportunity count. Drops, errors, and aborts therefore leave it below 100%. Scheduling state uses cyan, draining state uses yellow, failures use red, and drops, aborts, fallbacks, and nonzero rate limits use yellow. Zero alert counts are dimmed, and color is never the only indication of state. The downloaded metric is omitted in no-download mode. When stderr is not a TTY, the original ANSI-free tapbench progress: snapshots are emitted at most every five seconds, plus one at the schedule-to-drain transition. --no-progress suppresses both forms but not the final summary.

The final stdout report includes opportunity, submission, terminal, success, failure, HTTP, poll, WAIT, fallback, and byte counts; peak in-flight work; the separate throughput values; client-observed and UWS server-side latency percentiles; and a primary-failure breakdown. Its lifecycle section includes:

  phase WAIT requests / changes: 123 / 97
  WAIT unchanged / timeouts:     24 / 2
  fallback jobs / polls:         2 / 4
  WAIT disabled jobs:            0

Fallback reasons are also summarized by stable machine value when nonzero: request_timeout_too_short, nonactive_phase, http_rejected, malformed_job_document, immediate_unchanged, transport_timeout, and origin_suppressed. Result rows are omitted in no-download mode. Polling rate-limit totals continue to combine WAIT and /phase activity because both belong to the same lifecycle stage.

JSON output and exit status

--json-output records sanitized configuration (including phase_wait), UTC correlation timestamps, monotonic durations, query identities, aggregates, and one structured record for every opportunity, including local drops. Its schema id is tapbench.run/v5. The sanitized configuration records scheduling_mode as burst_once, paced_once, or fixed_duration, plus the exact opportunity_count for either once mode (null for fixed-duration mode). The run-level queries array records each query index, path, and SHA-256 digest; each opportunity record contains the selected query_index. The saved_results array reports each post-run destination, byte count, and any save error. Enums are strings and unavailable observations are null. Output is flushed and atomically replaces the destination, so a failed write does not leave a partially encoded report. Per-job records retain the three raw UWS timestamp strings alongside the derived queue and execution durations. They also expose the normalized UWS version; whether WAIT was disabled; physical WAIT, fallback, and timeout request counters; successfully parsed changed/unchanged WAIT counters; and the first durable fallback reason. Aggregate fields total these counters and provide a fallback-reason breakdown. Request counters include physical 429 attempts, while phase-change/unchanged counters include only successfully parsed non-429 responses.

Exit codes are:

On the first SIGINT or SIGTERM, scheduling stops immediately, active operations are cancelled, and already-created remote jobs receive bounded best-effort cleanup unless --keep-jobs was selected. Accumulated records are retained, the partial summary and requested JSON are produced, and the command exits 3. A second interruption may bypass cleanup and output.

Security

Basic authentication is applied explicitly to requests on the configured TAP origin. Redirects may cross origins so services can return results from object storage, but cross-origin requests never receive the TAP Authorization header. URLs containing userinfo or using unsupported schemes are rejected. Displayed and serialized URLs omit userinfo, query strings, and fragments, and job identifiers are redacted. Errors are bounded and scrubbed of credential values. Neither credentials, authorization headers, query text, nor complete job URLs belong in terminal output or JSON.

Never commit local credential files. In particular, .env.mppdb and backup variants are local secrets and are ignored by this repository.

APDB example queries

The queries in examples/ use table and column names from the published Rubin APDB schema:

• PREFIX.fast_lookup.sql returns at most ten current object summaries;
• PREFIX.selective_join.sql returns at most 25 reliable r-band detections joined to their current object positions;
• PREFIX.expensive_small_result.sql scans and aggregates source fluxes but returns only one row per band.

The mppdb.*.sql family uses unqualified development-server table names. The bq.*.sql family qualifies tables with the ppdb schema used by the integration deployment. All SQL files are included in source and wheel distributions under tapbench/examples/; tapbench examples stage accesses them through package resources without relying on an installation-specific filesystem path.

tapbench does not discover or rewrite schemas at runtime. Select the family whose qualification matches the target deployment, or stage and edit a local copy.

For a safe live check, first run each example once at a low rate with a single in-flight slot and normal cleanup. Then perform a short no-download benchmark:

uv run tapbench run \
  --url https://mppdb.juriclab.org \
  --rate 1 \
  --duration 3 \
  --max-inflight 1 \
  examples/mppdb.fast_lookup.sql

Do not use --keep-jobs for routine validation. Confirm the report contains three opportunities and no sensitive URL or credential material.

Development

The test suite uses a deterministic local fake TAP service; live service behavior is not a replacement for those tests. Run all checks with:

uv run ruff format --check .
uv run ruff check .
uv run mypy src
uv run pytest

Releasing

Package versions are derived from Git tags. Releases use clean tags of the form vX.Y.Z; for example, tag v0.1.0 produces package version 0.1.0. Untagged commits produce development versions and are never published.

Pushing a release tag builds and validates the wheel and source distribution, prints their complete file manifests, and saves them as workflow artifacts. It does not publish anything. Review the artifact names, hashes, and manifests, then create a draft GitHub Release for that tag.

Publishing the GitHub Release is the explicit production approval. It rebuilds and validates the tagged source, publishes the distributions to PyPI through Trusted Publishing, and verifies a fresh installation from PyPI. The publishing workflow stores no PyPI token and is restricted to the pypi environment.

Before the first release, create a pending Trusted Publisher on PyPI with project tapbench, owner mjuric, repository tapbench, workflow publish.yml, and environment pypi. Configure that GitHub environment to permit only tags matching v*.