Keeper of the limits
Scheduler is the application that, in the life-cycle of accepting, evaluating, and executing a build request, sits in the third position.
The first two positions are: Listener accepts a "request" (an incoming event
from GitHub). Gatekeeper evaluates, and configures (fetches .travis.yml
) the
request, and creates a build record, including at least one job per build in
the database.
Now, these jobs won't start automatically. They just sit in the database, and wait until they can be "queued" for the workers, so workers can pick them up, and start executing them. So, "queueing" a job, in this context, means: sending the job to the workers.
The app responsible for queueing jobs is travis-scheduler
. The current
implementation of the scheduler looks at all jobs that are in the created
state (i.e. they're "waiting to be queued"), groups them by their owner
(organization, user) and evaluates how many of them can be queued. Once done
evaluating jobs for all owners it starts over.
E.g. on travis-ci.org the default concurrency limit is 5, so the scheduler will make sure there aren't more than 5 jobs running at a time. If it finds 3 jobs already running for a given owner, and 4 jobs waiting to be queued, it will only queue 2 more jobs in order to respect the limit of 5.
The log output for each of these evaluation rounds looks like this:
user sven public capacity: total=3 running=1 selected=2
user sven boost capacity: total=2 running=0 selected=2
repo sven/repo: queueable=5 running=1 selected=4 waiting=1
user sven: queueable=5 running=1 selected=4 total_waiting=1 waiting_for_concurrency=1
The terminology used here can be confusing. The terms mean:
total
- number of concurrent jobs, provided by public capacity, plan, boost, etc.running
- number of jobs currently running, i.e. in the statequeued
,received
, orstarted
queueable
- number of jobs in the statecreated
selected
- number of queueable jobs that are being selected to be queued based on concurrency limitstotal_waiting
- total number of queueable jobs that have not been selected to be queuedwaiting_for_concurrency
- number of queueable jobs that have not been selected, and have not been found to be limited by repo settings, queue, or stages
In the example log output above, the owner has a capacity of 3 concurrent jobs
provided by public
capacity (line 1), and 2 jobs provided by boost
capacity
(line 2). The job selection finds 5 jobs to be queueable (i.e. in the state
created
), and 1 job to be running. As the total capacity is 5 jobs it can
select 4 jobs to be queued for the workers, leaving 1 job waiting.
See CONTRIBUTING for information on how to contribute to travis-scheduler.
See MIT-LICENSE.