Delivery Risk Radar

Which Apache Jira tickets run late, and why. Eight large open-source projects, 2007 to 2021, from a public snapshot of their Jira history.

What "late" means here. Almost no Apache ticket has a due date, so lateness is relative: a ticket is late when it took longer than 75% of comparable tickets (same project and type). It is a measure of being slow compared with peers, not of missing a promise. The data is a snapshot as of .

Where the time goes

Tickets spend most of their life waiting. The stage with active work is a few percent of lead time in most projects, and review takes a fifth to a third where it happens in Jira.

Compare stage shares within a group, not across. In the GitHub-review projects the review is invisible to Jira, so it shows up as waiting.

Waiting for a first review

In the four projects that review through Jira statuses, a typical ticket reaches review quickly. The slowest quarter is where the delay is.

Counting only delivered tickets understates lateness

Among delivered tickets the late rate seems to fall for recent tickets. That is censoring: the slow ones are not delivered yet. Counting open tickets that are already past their threshold as late removes the bias, and the model is trained on that version.

What the model relies on

At creation, the biggest signal is whether the ticket was assigned when it was filed. Tickets filed with an assignee are late far less often than unassigned ones (HIVE 34% against 72%). Who filed it and how it is described also matter. Priority barely separates late from on-time tickets.

Bars show how far each feature moves a ticket's score on average, from a test period the model never saw. These are associations: assigning a ticket does not make it faster. See Model quality and Method for how this was measured.