How we measured the figures in our release notes

Where the numbers in these notes come from

Every quantified claim published in a release note must be reproducible from a query recorded here, with the date it was run. A figure that cannot be re-derived is not published.

This record exists because the rule was broken: the September revision of metric-changes-august-2026 published "about 16%", "about 14%" and "99% of hierarchies" with no measurement behind them at all. Two could not be reproduced when checked, and one was wrong by roughly six times.

Duplicate task revisions — measured 2026-09-01, company 24472012

Run against production ClickHouse. clickup_tasks is insert-only between nightly compactions, so these are end-of-day figures and are lowest just after the 03:00 full sync.

-- Non-milestone row inflation: 4.28%  (93,335 rows / 89,505 tasks)
SELECT count(), uniqExact(task_id),
       round(100.0*(count()-uniqExact(task_id))/uniqExact(task_id), 2)
FROM clickup_tasks WHERE company_id = '24472012' AND is_milestone = 0;

-- Story-point inflation: 9.53%  (32,868 raw vs 30,009 collapsed)
SELECT round(100.0*(sum(sp_raw)-sum(sp_one))/sum(sp_one), 2) FROM (
  SELECT sum(storyPoints) AS sp_raw, argMax(storyPoints, synced_at) AS sp_one
  FROM clickup_tasks WHERE company_id = '24472012' GROUP BY company_id, task_id);
-- NOTE: this one includes milestones; the row figure above excludes them.
-- They are different populations and must not be compared directly.

-- Share of tasks with exactly one revision: 99.38%  (89,869 of 90,426)
-- Tasks carrying more than one: 557. Worst single task: 185 revisions.
SELECT countIf(n=1), countIf(n>1), round(100.0*countIf(n=1)/count(), 2), max(n)
FROM (SELECT count() AS n FROM clickup_tasks WHERE company_id = '24472012'
      GROUP BY company_id, task_id);
-- The maximum moves daily as revisions accumulate, so the note states it as
-- "more than 180" rather than pinning an integer that goes stale overnight.

Not published, because they could not be reproduced

"An August window inflates the accounting export by about 16%." The same shape of query over date_closed in August 2026 measures 2.7% (2,778 rows / 2,705 tasks). Six times smaller than the published figure. The claim was removed rather than restated, because the export's own query shape was never reproduced.
"Time Tracking's hours by about 14%." No measurement was ever run.
"About 99% of hierarchies are unaffected, a handful wrong by ten times or more." The 99.38% above is the share of tasks with one revision, which is a different population from descendant roll-ups. The roll-up amplification was never measured directly, so no figure for it is published.

Duplicate tracker seats — measured 2026-09-02, company 24472012

-- Names holding exactly two tracker seats: 35
SELECT count() FROM (
  SELECT name FROM users
  WHERE company_id = '24472012' AND clickup_user_id != 0
  GROUP BY name HAVING uniqExact(clickup_user_id) = 2);

Withdrawn. An earlier version of this file claimed "27 of those had both seats assigned to the same task" and asserted both figures reproduced. The query recorded for the first one referenced a column that does not exist (task_assignees_flat.name; the column is assignee), so it errored rather than returning anything, and the 27 figure had no query recorded at all and could not be reproduced when one was written. Both the assertion and the 27 were removed. This is the second time an unmeasured figure reached a published note, which is why the rule at the top of this record exists.

Deliberately not counted

The note describes the class of query that picks a task revision by close date rather than by sync order, and names the affected surfaces without giving a count. A count was published once — "eight queries" — and did not reproduce: depending on whether you count argMax expressions, distinct collapse sites, or executed queries after helper fan-out, the answer differs, and a single-line grep undercounts multi-line SQL. Naming the surfaces is what the reader can act on; the number was decoration that failed this record's rule.

Cost attribution — from the change's own record

This figure comes from the change's own record rather than from a query here: one engineer's July stage breakdown read 4,761 h against the 1,815 h attributable to them, 2.62x high. Scope is one engineer, one month, and the note says so — it must not be generalised to "the selected people".

← All releases