| Commit message (Collapse) | Author | Age | Files | Lines |
... | |
| | |
|
| |
| |
| |
| |
| |
| |
| |
| |
| | |
There are cases where a state machine handles an event but stays in the
same state. A callback is registered which filters messages further
before possibly taking action. There have been bugs caused by this
pattern being incorrectly implemented (where the callback is expected to
filter the message, but a transition takes place anyway). Hopefully a
consistent naming convention will make the pattern clearer.
|
| |
| |
| |
| |
| |
| | |
There is always one BuildController object per InitiatorConnection.
By coupling the objects slightly closer we can simplify some transitions
in BuildController.
|
| |
| |
| |
| |
| | |
This is similar to the issue fixed by commit
c38b77bed86acc8b90f253ce354f3ecf98e475e7.
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
Each BuildController instance sets itself up to receive all messages from
all workers (via the WorkerConnection instances). In the case of a build
failure, all BuildController objects would transition to 'None' state
(causing them all to be destroyed) on any WorkerBuildFailed message.
This meant that if any one build failed on a distbuild network:
- the user whose build actually failed would receive the error
messages correctly
- any concurrent users would see no further build messages from the
controller, making it look like their builds had hung.
Ctrl+C from the 'hung' users would still be correctly handled by the
controller, as their InitiatorConnection instance would still be alive.
|
| | |
|
| |
| |
| |
| | |
Makes it easier to see what they mean at a glance.
|
| | |
|
| |
| |
| |
| | |
This makes tracking multiple builds through one log file a bit easier.
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
New DistbuildSocket class that wraps socket.socket(), providing a
descriptive repr() handler showing where the socket is connected, and
providing a couple of helper methods for fetching local and remote
endpoint names.
This commit also adds a descriptive repr() handler to a few other
objects (mostly giving socket connection details).
|
| |
| |
| |
| |
| | |
We still log all messages sent to workers, which include the output of the serialise-artifact
code in full. There's no need for these status messages.
|
|/ |
|
| |
|
|
|
|
|
| |
We want to be able to transfer all
source artifacts in a single transaction
|
|
|
|
|
|
|
|
|
|
| |
Whenever the controller finds a source artifact
it wants to build, it changes its state to BUILDING.
We build all a chunk's source artifacts in one go.
So for any chunk artifact, we change the state of
all chunk artifacts that are built from the
same source to BUILDING
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Serialisation was simple when we only had 1 artifact per source.
However, to allow smaller systems, we need artifact splitting to produce
multiple artifacts per chunk source.
So now the new serialisation format has a separate list of artifacts
and sources, rather than the Source being generated from the artifact's
serialisation.
Python's id() function is used to encode the references between the
various Sources and Artifacts, these are replaced with a reference to
the new object after deserialisation.
Previously the cache-key was used, but this is no longer sufficient to
uniquely identify an Artifact.
The resultant build graph after deserialisation is a little different
to what went in: Strata end up with a different Source per Artifact,
so it _is_ a 1 to 1 mapping, as opposed to Chunks, where it's many to 1.
We serialise strata and chunks differently because stratum artifacts
from the same source can have different dependencies, for example
core-devel can have different dependencies to core-runtime.
Without intervention we would serialise core-devel and core-devel's
dependencies without including core-runtime's dependencies.
To solve this we've decided to encode stratum artifacts completely
indepedently: each stratum artifact has its own source. This is safe
because stratum artifacts can be constructed independently,
as opposed to Chunks where all the Artifacts for a Source
are produced together.
This is a little hacky in its current form, but it simplifies matters
later in distbuild with regards to how it handles expressing that
every Artifact that shares a Source is built together.
Arguably, this should be the output of producing the build graph
anyway, since it more helpfully represents which Artifacts are built
together than checking the morphology kind all the time, but more
assumptions need checking in morph before it's safe to make this
change across the whole of the morph codebase.
|
|
|
|
| |
Worker name is not sent in message
|
|
|