Dokimore Engineering Team
Builds and runs the platform, and writes up what was measured and how.
- Articles
- 14
- Topics
- 2
- Last updated
What this team covers
- Engineering
- Case Studies
Articles signed by this team
- EngineeringHow-to guide
How to Test File Uploads with Playwright: setInputFiles, the File Chooser and Proof the Server Got the Bytes
Playwright uploads a file with setInputFiles: from a path, from several paths, or from bytes built in the test with no file on disk. The test is finished when the server reports what it received and the test compares a hash. This guide runs all four shapes, the custom button that opens the file chooser, clearing a selection, and the failure path a size limit must produce, in Chromium, Firefox and WebKit.
- EngineeringHow-to guide
How to Test File Downloads with Playwright: the Download Event, the Race It Must Win, and the WebKit Case
Playwright reports a download as a download event with the suggested name, the URL and the bytes. The wait for it must be armed before the click, the saved file must be checked against what the server sent, and a server error must be caught as a failed download rather than a missing event. This guide runs an attachment link, a blob built in the page, a download that starts 1.5 seconds after the click, an inline PDF in a new tab and a failing export in Chromium, Firefox and WebKit, and reports where WebKit behaved differently.
- EngineeringHow-to guide
How to Debug a Failed Playwright Test: Read the Call Log, Open the Trace, Capture the Three Signals
A failed Playwright test tells you what it was waiting for; the call log under the timeout names the element, its state and what stood in the way. This guide reproduces the two most common failures, a selector that matches nothing and a button covered by an overlay, quotes the exact messages Playwright 1.62.1 printed in Chromium, Firefox and WebKit, shows what a trace archive holds, and lists the three signals (console, failed requests, 4xx and 5xx responses) a suite should capture on every run so the next failure explains itself.
- EngineeringOriginal research
XPath vs CSS Selectors for Test Automation: Measured on 10,000 Elements, and Which Ones Survive a Changed Page
CSS selectors are faster than XPath in every engine we measured, by a margin that does not matter in a test, and XPath can do two things CSS cannot: select by text and walk up to a parent. What decides a test’s maintenance cost is neither. It is whether the selector depends on position. When one row was inserted above the target, id, test-id, class and text selectors still found the same button in Chromium, Firefox and WebKit, and both positional forms, CSS nth-child and XPath li[N], pointed at the row above it.
- EngineeringHow-to guide
How to Test Login Flows and Authenticated Pages: Sign In Once, Save the State, Know When It Goes Stale
An authenticated test suite needs three things: one test that proves the login form works, a saved session that every other test starts from, and a way to notice when that session has expired. This guide runs the UI login, saves storageState, reuses it in a fresh context, logs in through the API with a CSRF token instead of the form, and shows what a stale session looks like, in Chromium, Firefox and WebKit, with the cookie attributes and the timings from the run.
- EngineeringOriginal research
Cross-Browser Testing with Playwright: What Chromium, Firefox and WebKit Reported About the Same Page
Cross-browser testing earns its cost only where engines differ, so this guide measures where they do. One fixture page was loaded in Playwright 1.62.1’s Chromium 151, Firefox 153 and WebKit 26.5 and asked thirty questions: user agent, viewport and screen, fonts, scrollbars, native inputs, CSS support, WebGL, automation tells. Twenty-two answers matched; eight differed, and two of those, a download that raised no event in WebKit and a positional selector 370 times slower there, broke or distorted tests in the sibling guides.
- EngineeringHow-to guide
API Schema Validation Testing: Catch the Breaking Change a Value Check Walks Past, with JSON Schema and OpenAPI
Most API tests assert a status and a few values, and a response can change shape underneath them without a single failure. Validating the response against a JSON Schema, or against the schema inside the API’s own OpenAPI document, catches the field that changed type, the field that went missing, the property that should not be there and the nested value that broke. This guide runs both against a fixture API before and after an "improved" release and prints the four errors Ajv produced.
- EngineeringOriginal research
Eight Days of Verified Crawler Logs on a New SaaS Site: OpenAI’s Bots Outread Googlebot
Original data, small and honest: the first eight days of crawler traffic on dokimore.com, verified by address rather than by name. OpenAI’s two crawlers made 232 requests to Google’s 143, GPTBot read 124 distinct pages on its first day, and eight page fetches were made for a person by an answer engine. How we verify a crawler instead of trusting its name, what Cloudflare’s network-wide numbers say the ratio should look like, and what a new site should take from it.
- EngineeringOriginal research
We Benchmarked Five Small Local Models on 250 Real Vulnerability Descriptions. Twice the Speed Cost 12 Points.
Our test intelligence runs on local models, so swapping a model is a product change that can move accuracy without anyone noticing. We ran qwen3.5:4b, qwen3:1.7b, granite3.3:2b, llama3.2:3b and llama3.2:1b through the same 250 labelled vulnerability descriptions, drawn from CISA’s public Vulnrichment corpus, and compared them pairwise. Accuracy ranged from 62.8% to 18.4%. The model twice as fast as the incumbent was not the more accurate, and the gap was statistically significant.
- EngineeringCase study
How We Made a React SPA Readable by AI Answer Engines Without Adding a Dependency
Google renders JavaScript in a deferred queue and says so; most other crawlers, as far as their vendors document it, do not render at all. Our marketing site is the same Vite and React application as the product, so we prerender every sitemap route at build time with Playwright, assert the head tags and the JSON-LD on each page, and name every crawler in robots.txt. Zero new packages. Here is the mechanism, the first weeks of crawl data after launch, and what Google says you do not need.
- Case StudiesCase study
How We Test Dokimore With Dokimore: Drift Tests, Quality Gates, and Dogfooding
Dokimore’s release pipeline runs on the platform we sell, and the suite that has corrected us most often is a set of drift tests: checks that the public pages, the pricing catalogue, the sitemap dates and the legal claims still match the code. This is what runs before an image ships, what it has caught, with dates, and what you can copy.
- EngineeringExplainer
Security Scanner False Positives: Evidence Instead of Alert Fatigue
Dynamic scanners produce long lists, and much of the list is a pattern that looked risky rather than a weakness anyone could use. Triage eats the time that should go to fixing, and after a few sprints nobody reads the report. Dokimore reports a finding as confirmed only after a verification replay reproduces it with a benign payload, and ships the request, the response and the proof with every one.
- EngineeringExplainer
The Regression You Can’t See: Why Microsecond API Timing Matters
When an API budget is 20 ms, a drift from 0.3 ms to 0.9 ms disappears on a whole-millisecond display until five such services sit in one request path, and a dashboard that bands its values hides 3 ms becoming 9 ms the same way. This article breaks a request into the phases the Resource Timing interface and k6 both measure, explains which phase to watch, and says exactly where Dokimore records microseconds and where it still displays whole milliseconds.
- EngineeringExplainer
Inside a Load Testing Engine: Apache JMeter’s Load Model Behind a Web UI
Apache JMeter models load well, and its plans are XML files edited in a desktop application and run from the command line. Dokimore’s performance engine keeps the thread groups, timers and assertions that load engineers think in, puts them behind a web UI where a plan is reviewable configuration, and records every sample in a histogram at microsecond resolution to report P50, P95, P99 and Apdex live. A profile set up by hand on the load screen exports as a k6 script, and there is no JMX import or export.
What its articles carry
- 14 of 14 list their outside sources, each with the date it was read.
- 14 of 14 say how their numbers were measured, under “How we measured this”.
- 14 of 14 answer the questions readers ask, one paragraph each.
- An article that changes after publication shows the date it was updated.