Job Lifecycle
This page describes the various states that a job or analysis may go through in its lifecycle. You can monitor executions on either the visual platform interface or the command-line interface.
For reference, we use the following terminology when discussing jobs and analyses:
execution: an analysis or job
root execution: the analysis or job that was created immediately as a result of a user making the API call to run a workflow, app, or applet
execution tree: the set of all jobs and/or analyses that were created as a result of running a root execution
analysis: an analysis is created when a workflow is run; it consists of some number of stages, each of which is represented by either another analysis (if running a workflow) or a job (if running an app or applet)
parent analysis: each analysis is the parent analysis to each of the jobs that were created to run its stages
job: a job is created when an app or applet is run, or when a job spawns another job; a job is always a unit of execution that is run on a worker in the cloud
origin job: the job created when an app or applet is run by either a user or an analysis; this job always executes the "main" entry point
master job: the job created when an app or applet is run by a user, job, or analysis; this job always executes the "main" entry point; all origin jobs are also master jobs
parent job: the job which creates another job via the /job/new API call
subjob or child job: a job created by a job via the /job/new API call; this job executes the entry point specified in the API call which created it and runs the same executable as its parent
job tree: a set of all jobs which share the same origin job
job-based object reference: a hash containing a job ID and an output field name; this hash is given in the input or output of a job and will be replaced with the specified job's output field by the DNAnexus platform once the specified job has transitioned to the done state
In the following example, we have a workflow that has two stages, one of which is an applet, and the other of which is an app.
If the workflow is run, it will generate an analysis with an attached workspace for storing intermediate output from its stages. Jobs are also created to run the two stages. These jobs in turn can spawn more jobs, either to run another function in the same executable or to run an executable. The blue labels indicate which jobs or analyses can be described using a particular term (as defined above).
Note that the subjob or child job of stage 1's origin job shares the same temporary workspace as its parent job. Any calls to run a new applet or app (using the API methods /applet-xxxx/run or /app-xxxx/run will launch a master job that has its own separate workspace, and (by default) no visibility into its parent job's workspace.
Every successful job goes through at least the following four states: 1. idle: initial state of every new job, regardless of what API call was made to create it. 2. runnable: the job's inputs are ready, and it is not waiting for any other job to finish or data object to finish closing. 3. running: the job has been assigned to and is being run on a worker in the cloud. 4. done: the job has completed, and it is not waiting for any subjob to finish or data object to finish closing. This is a terminal state, so no job will become a different state after transitioning to done.
Jobs may also pass through the following transitional states as part of more complicated execution patterns:
waiting_on_input (between idle and runnable): a job enters and stays in this state if at least one of the following is true:
it has an unresolved job-based object reference in its input
it has a data object input that cannot be cloned yet because it is not in the closed state or a linked hidden object is not in the closed state
it was created to wait on a list of jobs or data objects that must enter the done or closed states, respectively (see the
dependsOnfield of any API call that creates a job); linked hidden objects are implicitly included in this list
waiting_on_output (between running and done): a job enters and stays in this state if at least one of the following is true:
it has a descendant job that has not been moved to the done state
it has an unresolved job-based object reference in its output
it is an origin or master job which has a data object (or linked hidden data object) output in the closing state
There are two terminal job states other than the done state, terminated and failed, and a job can enter either of these states from any other state except another terminal state.
The terminated state is entered when a user has requested that the job (or another job that shares the same origin job) be terminated. For all terminated jobs, the failureReason in their describe hash will be set to "Terminated", and the failureMessage will indicate the user responsible for terminating the job. Only the user who launched the job or administrators of the job's project context can terminate the job.
Jobs can fail for a variety of reasons, and once a job fails, this triggers failure for all other jobs that share the same origin job. If an unrelated job (i.e. is not in the same job tree) has a job-based object reference or otherwise depends on a failed job, then it will also fail. For more information about errors that jobs can encounter, see the Error Information page.
Jobs can automatically restart upon certain types of failures, such as UnresponsiveWorker, ExecutionError, AppInternalError, JobTimeoutExceeded, which can be specified in the executionPolicy of an app(let) or workflow. If a job fails for a restartable reason, its failure propagates to its nearest master job and restarts that. The master job can be restarted a number of times, given in the executionPolicy, if it is out of retries, the entire job tree fails.
restartable: the job is a master job, and its job tree can be restarted.
restarted: the job has been been restarted at least once.
For unsuccessful jobs, there are a couple more states that jobs may enter between the running state and its eventual terminal state of terminated or failed; unsuccessful jobs starting in all other non-terminal states will be transitioned directly to the appropriate terminal state.
terminating: the transitional state when the worker in the cloud has begun terminating the job and tearing down the execution environment. Once the worker in the cloud has reported that it has terminated the job or otherwise becomes unresponsive, then the job will transition to its terminal state.
debug_hold: a job has been run with debugging options and has failed for an applicable reason, and is being held for debugging by the user. For more information about triggering this state, see the Connecting to Jobs page.
All analyses start in the state in_progress, and, like jobs, will end up in one of the terminal states done, failed, or terminated. The following diagram shows the state transition for all successful analyses.
If an analysis is unsuccessful, it may transition through one or more intermediate states before it reaches its terminal state:
partially_failed: this state indicates that one or more stages in the analysis have not finished successfully, and there is at least one stage which has not transitioned to a terminal state. In this state, some stages may have already finished successfully (and entered the done state), and the remaining stages will also be allowed to finish successfully if they can.
terminating: an analysis may enter this state either via an API call where a user has terminated the analysis, or there is some failure condition under which the analysis is terminating any remaining stages. This may happen if the
executionPolicyfor the analysis (or a stage of an analysis) had theonNonRestartableFailurevalue set to "failAllStages".
In general, compute and data storage costs due to jobs that end up failing because of user error (e.g. InputError, OutputError) and terminated jobs are still charged to the project in which the jobs were run. For internal errors of the DNAnexus platform, such costs will not be billed.
The costs for each stage in an analysis is determined independently. If the first stage finishes successfully while a second stage fails for a system error, the first stage will still be billed, and the second will not.
