Documented size, row-count, duration, and resolution ladders for measuring parser, renderer, converter, and upload performance without relying on private production data.
A 50,000-row transactions dataset — an in-repo 'large' fixture for testing streaming CSV parsers, import performance, and pagination. Deterministic (fixed seed).
A CSV with 1000 columns and a handful of rows — a stress test for parsers, spreadsheet importers, and databases with column limits or per-row buffering assumptions. Deterministic (fixed seed).
A deliberately bounded compression-ratio fixture: 4474 bytes on disk expanding to exactly 4 MiB of zeros, roughly 937:1. It is not a decompression bomb - there is no nesting and no recursion, and the expanded size is a fixed 4 MiB - so it exercises a ratio guard or a streaming extractor's memory ceiling without being dangerous to open.
The same three members followed by the end-of-archive blocks and then zero padding out to a whole 10240-byte record - tar's historical blocking factor of 20, still the default. The archive is 10240 bytes for 616 bytes of content, which is what makes tiny tarballs look so wasteful before compression.
Position, normal and UV packed into ONE bufferView with byteStride 32 and three accessors at offsets 0, 12 and 24 — the GPU-friendly layout, and the one a naive parser mis-reads because it assumes each accessor owns its view. Twinned here with a tightly-packed version of the same sphere.
The same sphere with every attribute in its own bufferView and no byteStride — the layout most exporters emit and most parsers assume. The control against which the interleaved twin in this group is diffed.
An unquantised sphere: 32-bit floats for both positions and normals, 24 bytes of attribute data per vertex. The uncompressed reference for the quantized and gzip twins in this group.
The same sphere with positions as normalised 16-bit and normals as normalised 8-bit integers under KHR_mesh_quantization — 66.6% of the float32 twin's bytes. The extension is in extensionsRequired, so a loader without support must refuse the file rather than render garbage.
The float32 GLB from this group as a deterministic gzip stream (no embedded filename, mtime 0) — what a CDN actually sends when a viewer advertises `Accept-Encoding: gzip`. It compresses to 29.2% of its source against 55.6% for the quantized twin, which is the point of shipping both: quantisation halves the file on disk and then gives most of that back over the wire.
The quantized GLB from this group as a deterministic gzip stream (no embedded filename, mtime 0) — what a CDN actually sends when a viewer advertises `Accept-Encoding: gzip`. It compresses to 55.6% of its source against 29.2% for the float32 twin, which is the point of shipping both: quantisation halves the file on disk and then gives most of that back over the wire.
The same unit sphere at 16,128 triangles — level 0 of a four-step ladder that spans 202x from top to bottom. Because every level is generated analytically from the same surface rather than decimated, the ladder is an exact reference: any deviation you measure is your tool's, not the fixture's.
The same unit sphere at 3,968 triangles — level 1 of a four-step ladder that spans 202x from top to bottom. Because every level is generated analytically from the same surface rather than decimated, the ladder is an exact reference: any deviation you measure is your tool's, not the fixture's.
The same unit sphere at 960 triangles — level 2 of a four-step ladder that spans 202x from top to bottom. Because every level is generated analytically from the same surface rather than decimated, the ladder is an exact reference: any deviation you measure is your tool's, not the fixture's.
The same unit sphere at 80 triangles — level 3 of a four-step ladder that spans 202x from top to bottom. Because every level is generated analytically from the same surface rather than decimated, the ladder is an exact reference: any deviation you measure is your tool's, not the fixture's.
An 8x8 field of columns built from a single 12-triangle mesh referenced by 64 nodes with seeded random heights. Core-glTF instancing needs no extension: the file is tiny, and the interesting question is whether your renderer batches the 64 draws or issues them one by one.
Thirty-two copies of one cube expressed as instance attribute accessors on a single node under EXT_mesh_gpu_instancing — the extension that moves instancing off the scene graph and onto the GPU. A viewer without support shows exactly one cube at the origin, which is the tell.
The two-axis font with its width axis pinned to Condensed and its weight axis left live — the shipping shape for a site that wants one width but the whole weight range. For testing that a loader reads the reduced fvar rather than the original axis list. Original typeface, free to use.
The complete face the rest of the subset set is cut from: the whole glyph repertoire with ligatures, kerning, contextual alternates, small caps, stylistic sets, figures, fractions, marks and localised forms intact. Measure it, then measure the cuts. Original typeface, free to use.
The full face as WOFF2 — the honest baseline for a subsetting benchmark, since comparing a subsetted WOFF2 against an unsubsetted TTF measures Brotli rather than the subsetter. Original typeface, free to use.
The full face subset to U+0020-U+007E (printable ASCII) — the range a Latin-only page actually needs. Layout features are kept, so lookups whose glyphs were removed are the ones you can watch disappear. Original typeface, free to use.
The latin basic subset in the container a site would serve, so a size comparison against the full WOFF2 measures the subset rather than the compressor. Original typeface, free to use.
The full face subset to the 8 codepoints of the string NOVUS 07 — the extreme case a headline-only web font gets cut down to. Layout features are kept, so lookups whose glyphs were removed are the ones you can watch disappear. Original typeface, free to use.
The headline subset in the container a site would serve, so a size comparison against the full WOFF2 measures the subset rather than the compressor. Original typeface, free to use.
The full face subset to U+0030-U+0039 (digits only) — a numeric-display cut, the shape a dashboard font ends up. Layout features are kept, so lookups whose glyphs were removed are the ones you can watch disappear. Original typeface, free to use.
The digits subset in the container a site would serve, so a size comparison against the full WOFF2 measures the subset rather than the compressor. Original typeface, free to use.
A subset cut to just the characters the kerning, ligature, small-cap and figure probes need, keeping every OpenType layout feature whose glyphs survived. Half of a pair whose only difference is the layout tables — the pair that shows what a default subsetter setting silently costs you. Original typeface, free to use.
The same probe-character subset with every OpenType feature stripped — what many web-font pipelines produce by default. GSUB and GPOS survive as empty shells, so a tool that merely checks whether the tables exist reports the font as fully featured while nothing ligates, kerns, or turns into small caps. Original typeface, free to use.
The uncompressed TrueType (glyf) source for this container set. Every WOFF and WOFF2 file in the group is packed from exactly these bytes, so a compression comparison is measuring the container and nothing else. Original typeface, free to use.
The same TrueType (glyf) face in a WOFF container, 49 percent of the raw sfnt size. Compare it against the other three containers in this group to see how much of a web font's weight is the format and how much is the design. Original typeface, free to use.
The same TrueType (glyf) face in a WOFF2 container, 33 percent of the raw sfnt size. Compare it against the other three containers in this group to see how much of a web font's weight is the format and how much is the design. Original typeface, free to use.
The uncompressed CFF (Type 2) source for this container set. Every WOFF and WOFF2 file in the group is packed from exactly these bytes, so a compression comparison is measuring the container and nothing else. Original typeface, free to use.
The same CFF (Type 2) face in a WOFF container, 55 percent of the raw sfnt size. Compare it against the other three containers in this group to see how much of a web font's weight is the format and how much is the design. Original typeface, free to use.
The same CFF (Type 2) face in a WOFF2 container, 45 percent of the raw sfnt size. Compare it against the other three containers in this group to see how much of a web font's weight is the format and how much is the design. Original typeface, free to use.
A TrueType Collection holding three faces of one family with identical tables stored once and referenced by offset from each face's table directory. For testing face-index selection, collection-aware loaders, and extractors that must not assume each face owns its tables. Original typeface, free to use.
The same three faces in a collection that stores every table separately, so each face's directory points at its own copy. Identical rendering, a bigger file — the reference that makes the shared-table saving measurable rather than assumed. Original typeface, free to use.
A traffic-advice document opting a fictional origin into private prefetch proxy traffic at full fraction while disallowing every other agent. For testing prefetch-proxy discovery and the JSON-array (not object) top level this format uses.
240 cases across six classes, three of them failing. Big enough to time a parser and to check that a report UI paginates or virtualises rather than rendering every row, but still small enough to read.
A five-request page load — document, stylesheet, script, image and favicon — across two fictional hosts, with full per-phase timings and an onLoad page timing. Every entry's time equals the sum of its non-negative timing phases, so a waterfall renderer can be checked arithmetically rather than by eye.
A four-call REST session against a fictional API host, with JSON request bodies in postData, an Authorization header carrying an obvious SAMPLE placeholder token, and a 204 response that has no body at all. The right shape for testing a HAR-to-mock-server or HAR-to-code converter.
An http URL that redirects to https and then to a query-carrying cart page, with redirectURL populated on both hops. Tools that count entries as page views report three here when the user saw one page.
Five failing requests covering authentication, authorisation, a missing resource, rate limiting with a Retry-After header, and a server error that returns HTML instead of the JSON the client asked for. The last one is the case that breaks clients which parse by status code rather than by content type.
A full session-cookie lifecycle: set at login, echoed on the next request, then expired at logout with Max-Age=0. Both the header and the parsed cookies array are present, and every value — including the form password field — is an obvious SAMPLE placeholder rather than a credential.
Four API calls in which the wait phase — time to first byte — accounts for almost all of the elapsed time, the signature of a server-bound rather than network-bound page. Use it to check that a performance report attributes the cost to the server instead of blaming connection setup.
One form submission whose postData carries both the parsed params array and the raw urlencoded text — and the two must agree. The notes field contains a comma and an ampersand, so a decoder that splits before unescaping produces seven fields instead of five.
A two-part multipart/form-data upload: a text file part with a fileName and a plain field part. The body uses CRLF between MIME parts because RFC 7578 requires it, even though the surrounding HAR document is LF-terminated — a distinction that matters when a replay tool rewrites the body.
A PNG response whose body is carried as base64 with content.encoding set accordingly. content.size is the DECODED byte count, not the length of the base64 string, which is the field readers most often get wrong when computing transfer weight.
A gzip-compressed JSON response where content.size is the uncompressed length, bodySize is the bytes on the wire, and content.compression is the saving between them. It also carries the non-standard _transferSize field browsers add, which readers must tolerate rather than reject.
Two navigations in one capture, with entries bound to their page through pageref. The second page's onLoad is -1, the HAR convention for a timing that never happened, which naive code averages in as zero and reports as an impossibly fast load.
A conditional request answered 304 with full beforeRequest and afterRequest cache state, next to a first-time fetch whose beforeRequest is null. The null is the documented way to say "not in the cache", and readers that treat it as a missing object rather than an explicit absence lose the distinction.
An HTTP/2 entry with lowercase pseudo-headers, an httpVersion of "h2" rather than "HTTP/2.0", an ssl phase inside connect, and three underscore-prefixed custom fields. Version strings are not standardised across writers, so anything matching on "HTTP/1.1" misses this entry entirely.
One request whose query string encodes a space as %20 in one parameter and as + in another, includes a parameter with an empty value, and carries an Accept-Language header with quality values. The parsed queryString array is the answer key for whatever a URL parser produces from the raw URL.
A load-test run summary in the shape a threshold-checking CI step consumes: totals, per-scenario percentiles, and three named thresholds each carrying its limit, the observed value and a pass flag. Publishing the limit next to the observation is what lets a gate be re-evaluated without rerunning the test.
Three hundred individual request records with timestamp, scenario, status and duration — the raw log a percentile calculation has to be computed from rather than read off. A handful of 4xx and 5xx responses are mixed in so an error-rate calculation has something to find.
The percentile table a performance report renders as a chart, with mean, p50, p95, p99 and max for each of the four scenarios. The same numbers appear inside the JSON summary in this group, so a chart built from either source can be diffed against the other.
A budget file pairing per-path resource-size and timing limits with API latency budgets, including a glob path that must be matched rather than compared for equality. It is the configuration side of the load-test results in this group: the limits a run is judged against.
A single-service trace 48 levels deep — a recursive GraphQL-style resolver where every span is the child of the one before it. A depth fixture for waterfall renderers, tree-layout code and any parser that recurses over parentage.
One root span with 60 overlapping children — a scatter-gather query across database shards. The counterpart to the deep-chain fixture: shallow but very wide, which is what breaks waterfall layouts that assume spans do not overlap in time.
A single metric family blown up to 2,000 series by two unbounded labels — customer_id and session_id — which is the most common way a metrics bill and a TSDB both fall over. A load fixture for scrape parsing and a worked example for cardinality-limit rules.
Every record carries five fields that are unique to it — session, customer, device, pod name and container ID. Harmless as log text, ruinous the moment they are promoted to indexed labels, which is the mistake this fixture exists to make measurable.
A single record of about 75 KB — a request body accidentally logged — sitting between two ordinary ones. Most shippers cap a line at 16 or 64 KB and either split it into invalid JSON fragments or drop the records around it; this file makes which one happen visible.
A CPU profile in the folded-stack format every flamegraph tool reads: one line per unique stack, frames separated by semicolons, then a space and a sample count. Twelve stacks up to seven frames deep, with the runtime's GC and scheduler work included as real profiles have.
The same workload after a change: JSON decoding costs four times as much and gains a new reflect frame, while cache lookups get cheaper. Paired with the baseline so a differential flamegraph can be scored against a known regression and a known improvement.
An allocation profile in the same folded syntax as the CPU profiles, but weighted in bytes rather than samples. The format carries no unit, so a viewer that assumes samples renders 65 MB of allocation as 68 million samples — which is what this fixture is for.
A 30-second wall-clock profile with the thread name as the root frame, so time blocked in reads is visible instead of invisible. Off-CPU time dominates — the shape that tells you a service is waiting, not computing, which a CPU profile alone cannot show.
A recursive-descent parser profiled mid-recursion: stacks up to 204 frames deep where one frame repeats 200 times, on lines over 3,000 characters. Flamegraph renderers with a fixed depth limit truncate this silently, and line-length limits split it into nonsense.
A heap profile measuring bytes still live at the sample point, not bytes ever allocated — which is why its shape differs from the allocation profile of the same process. 620 MB retained across five sites, with an unbounded session cache at the top.
A speedscope sampled profile: one shared frame table with source locations, then stacks expressed as arrays of frame indices with per-sample weights in nanoseconds. The index indirection is the whole design, and resolving it wrong produces a plausible but wrong flamegraph.
The evented half of the speedscope format — explicit O and C events that must nest like brackets, rather than sampled stacks. Tracing profilers emit this shape, and an unbalanced or out-of-order close is the corruption a loader has to detect.
The .cpuprofile shape V8 emits: a node tree with call frames, a samples array of node IDs and a parallel timeDeltas array, plus the synthesized (idle), (program) and (garbage collector) nodes. Counting those synthetic frames as application time is the classic misreading.
The text report pprof prints for the same profile as the folded fixtures: a metadata header, then flat and cumulative columns where frames that are never a leaf carry zero flat time and near-total cumulative time. Column-aligned output with a bare 0 rather than 0s, which is where naive column parsing fails.
Line-level profile output: two routines with per-source-line flat and cumulative cost, a dot for lines that cost nothing, and units that switch between seconds and milliseconds within one column. Source lines keep their real tabs, which most column parsers do not survive.
Raw perf script output — a header line per sample followed by tab-indented call-chain frames with addresses, symbols, offsets and DSO paths, samples separated by blank lines. The input side of stackcollapse, including unresolved frames a symbolizer could not name.
The index that lets a reader seek directly to record N of the .shp instead of walking the file. Its header duplicates the .shp header exactly, and a mismatch between the two is the most common way a hand-edited shapefile goes wrong.
Unlike the point index, these entries have different lengths because each PolyLine holds a different number of vertices. That makes it the useful case for testing seek logic — a reader that assumes a fixed stride reads the first record correctly and then garbage.
The parcel index, whose second entry is markedly longer than the others because that record carries two rings. Use it together with the roads index to check that seek logic handles multi-part records as well as multi-record files.
A small raster tile pyramid covering the invented territory, one flat colour per zoom level so a wrongly addressed tile is obvious on sight. The row axis follows the TMS convention, which is flipped relative to the XYZ scheme web maps use — the single most common reason an MBTiles layer renders upside down.
We use Google Analytics and show ads via Adsterra. Non-essential cookies and ad scripts run only after you allow the matching categories. See our cookie policy.
