boinc/doc/backend_state.html

<title>Back end state transitions</title>
<h2>Back end state transitions</h2>

<p>
The processing of workunits and results involves
several independent activities.
To keep track of these activities,
workunit and result database records have several "state" fields,
and their processing can be viewed as the combination
of several finite-state machines.

<p>
A workunit has the following state fields:
<ul>
<li>
<b>delay_bound</b>.
upper bound for the interval between sending this WU to a host
and getting the result
Should be several times the execution time on an average host.
If it's exceeded, the server "gives up" on the result
and may delete its input files.
If the result is returned later,
it will still be validated and credited.

<li>
<b>canonical_resultid</b>.
<li>
<b>timeout_check_time</b>.

<li>
<b>file_delete_state</b>:
Initially INIT.
When the main state transitions to either DONE or ERROR,
it transitions to READY,
indicating that input files can be deleted.
When file deletion is completed (by file_deleter)
it transitions to DONE.

<li>
<b>assimilate_state</b>:
Initially INIT.
When the main state transitions to either DONE or ERROR,
it transitions to READY,
indicating that the workunit can be assimilated.
When assimplateion is completed (by assimilator)
it transitions to DONE.

<li>
<b>need_validate</b>:
A boolean, true whenever
the workunit has a result whose validate state is NEED_CHECK.
The validate program sets it back to false.
<li>
<b>error_mask</b>.
bit mask for error conditions

</ul>
Invariants:
<ul>
<li> eventually either canonical_resultid or error_mask is set
<li> eventually timeout_check_time=0
<li> WUs are eventually assimilated
<li> input files are eventually deleted,
but only when all results have state=OVER
            (since may need to validate results that arrive after assimilation)
        and wu.assimilate_state = DONE
            (since project may want to do something with WU in error case)


</ul>


<p>

A result has the following state fields:
<ul>
<li> <b>report_deadline</b>:
        give up on result (and possibly delete input files)
        if don't get reply by this time.
        Assignment: when send result; now + WU.delay_bound

<li> <b>server_state</b>:
UNSENT, IN_PROGRESS, OVER.
Initially UNSENT.
Becomes IN_PROGRESS when the result has been sent to a client.
Becomes OVER if we get a host reply,
or the result times out, or we decide not to send it.
<li> <b>outcome</b>:
SUCCESS, COULDNT_SEND, CLIENT_ERROR, NO_REPLY, DIDNT_NEED.
Defined if server_state = OVER.
<li>
<b>client_state</b>:
Records the client state (upload, process, or download)
where an error occurred.
Defined if outcome is CLIENT_ERROR.
<li>
<b>file_delete_state</b>:
INIT, READY, DONE.
<li>
<b>validate_state</b>:
INITIAL, VALID, INVALID.
When a canonical result has been found for the workunit,
becomes either VALID or INVALID.
</ul>

Invariants:
<ul>
<li> results eventually have server_state = OVER.
<li> output files are eventually deleted.
Non-canonical results can be deleted as soon as the WU is assimilated.
Canonical results can be deleted only when all results have server_state=OVER.
If a result reply arrives after its timeout,
the output files can be immediately deleted.
How do we delete output files that arrive REALLY late?
(e.g. uploaded after all results have timed out, and never reported)?
Let X = create time of oldest unassimilated WU.
Any output files created before X can be deleted.
</ul>
<h3>A note on scheduling</h3>
<p>
    - when is it feasible to send a result to a host?
        Request msg should include X = amount of work currently queued.
        TODO: include % time active in calculation??
        Decision for each WU:
            is X + time for WUs sent so far < delay_bound?
    - When is a result declared "unsendable"?
        Not a good idea to do on the basis of time;
        do it only if a result is flushed from FIFO (see below)

<h3>State transitions</h3>
<pre>

fields of "result" table:

server_state
    UNSENT
        (on creation)
    IN_PROGRESS
        from UNSENT
            scheduler: when send
    OVER
        from IN_PROGRESS
            scheduler: get reply from host
            timeout_check: now > report_deadline
        from UNSENT
            validate: got canonical result for this WU and server_state=UNSENT
            timeout_check: WU has error

file_delete_state
    INIT
        (on creation)
    READY
        from INIT:
            scheduler: got reply and server_state = OVER
            timeout_check: all results are OVER or report_deadline has passed
            assimilator: all results are OVER or result is not canonical
        from DONE:
            scheduler: got reply and server_state = OVER
    DONE
        from READY
            file_deleter: tried to delete files

validate_state
    INIT
    VALID
        from INIT:
            validate: outcome = SUCCESS and matched canonical result
    INVALID
        from INIT:
            scheduler: got reply, client error
            validate: didn't match canonical result

-------------
fields of "workunit" table

need_validate
    FALSE
        (on creation)
        from TRUE:
            validate: done checking
    TRUE
        from FALSE:
            scheduler: got reply w/ client_state = DONE (i.e. no error)

file_delete_state
    INIT
        (on creation)
    READY
        timeout_check: all results haver server_state=OVER
            and wu.assimilate_state = DONE
        assimilate:
            all results have server_state = OVER
            (and wu.assimilate_state = DONE)
    DONE

assimilate_state
    INIT
        (on creation)
    READY
        from INIT:
            timeout_check: WU has error
            validate: found canonical result
    DONE
        from READY:
            assimilator: done

error_mask
    COULDNT_SEND
        timeout_check: some result has outcome COULDNT_SEND
    TOO_MANY_ERROR_RESULTS
        timeout_check: too many error results
    TOO_MANY_RESULTS
        timeout_check: too many results

timeout_check_time:
    nonzero
        (on creation)
    zero
        timeout_check: all results are OVER and validate_state = DONE
</pre>