fix: #8624
1. Add a new flag `--force` for `queue kill`
2. Make `SIGINT` to be the default option and `SIGKILL` to be with
`--force`
3. Add tests for `queue kill`
4. Bump dvc-task into 0.1.9
fix: #8612
For checkpoint experiment in some case users might want to give it a early
stopping to reduce variance. But currently if we interrupt/kill the
experiment it will be marked as failed, and all of the completed
checkpoints will be removed as we cleanup all the running directly after
the process failed.
1. We raise CheckpointKilledError intead of StageCmdFailed error if at
least one checkpoint had been commited.
2. Temp_dir executor will continue collecting data if the Checkpoint
stage was interrupted.
3. Raise warnings if a checkpoint stage was incomplete and the other
stages were not forwarded.
4. Add a new functional test for this
fix: #8702
1. replace simple `proc.kill` in `queue stop --kill` with the same logic as in
`queue kill` to make it possible to set false status `running` tasks
as faileed.
2. Modified the tests.
fix: #8615
1. This is because, at the end of the tasks,
the temp running directly will be deleted
1. Only fetch result during running (not during the result collection)
2. Bump into scmrepo 0.1.4 in which we wrapped `KeyError` into
`SCMError`.
fix: #8448
In our current temp dir executor, we use will first set EXEC_* branches
to revisions retrieved from stash entry. And push them to remote, but
because the push operation is very slow, our locking operation will make
the whole repository unusable.
1. Now we set revisions from stash entry to some temp branches and then
push them to remote.
fix#7693
The problem was when we gathered experiments, and at the same time,
the executor are creating new commits. It will make the checkpoint
top/head return by `get_running_exps`/`get_ref`/`branch_revs` not the
same one. And cause the problem described in #7693
To solve this we did
1. move the `get_rev` to the beginning of `exp show` to make sure the
return results in `branch_revs`
2. Add a new function to move all of the properties that only belong to
`head` of the checkpoint from the wrong place to the top.
3. Add a new unit test for it.
4. Force fetch_running_exp to return data.
5. fix a bug in tmp_dir_executor.
fix: #8461fix: #8473
For now, for those inactive but uncompleted tasks, they still shown as
Running in `exp show` and `task queue`, we cannot remove them because
they are regarded as running, we cann't kill them because their are no
real process behind them.
1. Now we can examine these false running tasks and mark them as failure
in `queue kill` command. #8461
2. Better msg handle for initliaizing tasks in `queue kill`#8473
3. Add new unit tests for both cases.
4. bump into new dvc-task 0.1.6
fix: #8526
1. The current `get_names` will crash if an experiment branch had
already been given a new branch name to it.
1. Solve the crash when we getting the name of a newly fixed experiment
branch.
2. Add a new unit test to it.
fix: #8350
We shouldn't add locks to exp show to prevent race conditions
as this command is calling continuously by the extension.
So we make a better way to handle all kinds of reference failures.
1. Drop all based experiments if got a SCMError in baseline.
2. Drop all checkpoint experiments if any commit got a SCMError in
one of checkpoint commit.
3. Drop the entire experiments if got a SCMError.
4. Add tests for each of the conditions above.
fix: #8349
`tasks status turned to queued for 1 second before turned into success.`
is because in previous the status is using the output of
`get_running_exps` in `repo.expereiments` and when entering the
collect result progress, the `get_running_exps` will mark the tasks to
be not running because the `info.result` is not None. Here we use the
output of the celery queue as the universal standard for the tasks
status.
1. Use celery status as the exp show status
fix: #8451
1. Change the call of scm.describe from individual revision to a
collection of them. Accelerate the run speed of `exp show`
2. Bump scmrepo to 0.1.2
fix: #7986
1. Add two new flags `--hide-queued` and `--hide-failed` to `exp show`
2. Allow `exp show` to show failed experiments.
3. Add unit test for the failed experiments shown.
4. Add name support for failed exp
5. Add error msg to the `exp show` output
Currently infofile dump was spread in both executor and queue classes.
put them all into executor to make them more managable. Move all ending
dump operations to the cleanup function, because the collect_result progress
might failed because of git operations, in which can guarantee them to
be run in a `finally` scope.
1. Move some basic test script from function tests to unit test.
2. Add success/failed tests for the status change of `tempdir`, `celery`, `workspace` running case.
Reason:
1. Need some more granular control over initialization progress of
executor.
What done:
1. Seperate `init_git` and `init_cache` progress out from setup_executor
2. Move `set_ref` into `init_git`
This is a flaky, the queue tasks status might remain "Running" even if
it had already been completed, beside sometimes I also noticed it raise
a pytest fixture error.
fix: #8014
> ERROR: Invalid experiment '{entry.stash_rev[:7]}'.
This happens when the queue task failed from a scm error.
{"exc_type": "GitMergeError", "exc_message": ["Cannot fast-forward HEAD
to '05100047a341f2fa4a02421289d48f84d8c45e86'"], "exc_module": "dvc.scm"}.
And the failed task neither create a infofile nor create a fail_stash.
1. Didn't raise DvcException if no infofile found for failed tasks.
(success tasks still raise it)
2. Add a new unit test for this.
fix: #8018fix: #8071
1. Add one line to make `exp remove` can remove failed tasks, and will
also delete associated successful tasks
2. Add a new unit test for this behaviour.
3. Solve the unstable test_queue problem.
fix: #8031
Currently when we use tmp directory to run a experiment, we set
`EXEC_HEAD` , `EXEC_BASELINE` and `EXEC_MERGE` heads locally but forget
to remove it. This cause some problem when we use `exp show`
1. Clean `EXEC_HEAD` , `EXEC_BASELINE` and `EXEC_MERGE` after we pushed
them to tmp_dir
2. Add a unit test for it.
Co-authored-by:
Followed from #7592
1. Seperate remove method to a new file.
2. Add queued, failed, processed flags to remove.
3. Add new unit tests for queue status --queue/fail/success
4. Implement methods to remove done tasks.
5. add some new unit and functional test for celery_queue.remove
6. bump into dvc-task version 0.0.13
fix: #7950
Currently queue worker status are not show in `queue status` makes it
hard to tell if the worker is still living.
1. add a new method `active_worker` to return the currently running
worker node name.
2. give every worker a unique name.
3. `queue status` will show how much worker running at present.
Currently, all of our celery workers are called `dvc-exp-worker` and
they will write to the same log path.
We give every single worker a unique name and seperate they logs if
under a debug log level
fix: #7840
Now queue remove can not accept short rev as name.
1. Refactor the queue entry match in `queue remove`
2. Better output if no queue task found for name.
3. short rev on kill can only be work for rev with lens longer than 4
fix: #7590
1. Add two new command `queue status` and `queue attach`
2. Add unit test for the CLI
3. Make `Running` exps can be found by `get_queue_entry_by_names`
1. Add some more unit test for `celery_queue.kill`
2. Implement `kill` argument for `celery_queue.shutdown`
3. Use dataclass's internal methods instead of `__eq__`
fix: #7591
1. Add a new command `dvc queue`
2. Add two sub-command `dvc queue remove` and `dvc queue kill`
3. Add a unit test to test them
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
fix: #7587
1. Add implement kill method for local queue class
2. Add a test to make sure the following job will be success after the
original job was killed.
3. Some refactoring work on `exp remove`
1. Implement `--time` to `exp gc`
2. add a unit test for it.
3. Rename `--time` to `--date` and change the warning msg.
4. To get a better msg for `gc` and `exp gc`
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
Co-authored-by: Dave Berenbaum <[email protected]>
1. Add argument `commit_time` to `gc`/`brancher` and `iter_revs`
2. Add unit test for iter_revs.
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
fix#678
1. Add a new flag `--time` to `dvc gc`.
2. Modify the unit test of `dvc gc`
3. Fix an argument checking error and update the error msg.
Co-authored-by: Dave Berenbaum <[email protected]>
fix: #7448
Current we skip on those experiments which already existed on both sides, but
do not exclude them from the result list. And if we have diverged experiment,
pulling/pushing them will stop the current progress and raise an exception, but
didn't give any information on the condition of other experiments.
1. Change the exp pull and push to match the new API in `scmrepo`.
2. Handle sync status after `exp pull/push` finished.
3. Bump scmrepo to 0.0.19
Co-authored-by:
fix: #7449
`exp remove`: should show the experiments that were removed
1. experiment.remove now returns the experiment list been removed.
2. modify the functional test to these the returned list.
3. use ui to show the removed list.
Co-authored-by: Dave Berenbaum <[email protected]>
fix: #7155
1. rename flags `-A/--all-commits` in `exp remove`
2. add new flag `-n/--num` in `exp remove`
3. unify the revision collection in `exp remove`
4. add unit and func tests for `exp remove`
Co-authored-by: Jorge Orpinel <[email protected]>
fix: #7154
1. rename flags `-A/--all-commits` in the `exp pull/push`
2. add new flag "--rev" and "--num" in the `exp pull/push`
3. Unify the collection of revs in `exp push/pull`
4. add unit and func tests for `exp push/pull`
fix: #7153
Unify the flags in `exp list` to `exp show`
1. Add new `-A`, `-n` flags to `exp list` command.
2. Add unit test for the flag parse.
3. unify the revs retrive logic to `exp show`.
4. Add functional test for `exp list`
5. Update the documents.
Co-authored-by: Jorge Orpinel <[email protected]>
fix: #7329
In a previous PR, we introduce a bug which use `HEAD` as the workspace
commit, while if there is a checkpoint is running, the `HEAD` is
actually changing all the time, we should use
`refs/exps/exec/EXEC_BASELINE` instead.
1. add `fix_exp_head` transfer in `dvc/scm.py::resolve_rev`
2. remove fix_exp_head from `diff`, `metrics diff`, `params diff` and
`plots diff`
3. modify a functional test to simulate the running conditions
fix: #7196
New feature: now `dvc exp pull` can accept a list of experiment names as input.
2. add support to list input for the `exp pull`
3. add function tests for the multi exp pull
4. add unit test for the arg parse
fix: #7196
New feature: now `dvc exp push` can accept a list of experiment names as input.
1. refactor of experiment push
2. add support to list input for the `exp push`
3. add function tests for the multi exp push
4. add unit test for the arg parse
fix: #7152
1. Add `--rev` flag to `dvc exp show`.
2. Add -1 support for `--num` flag`.
3. Extract revision found logic to `utils`.
4. Brancher use this revision found logic.
5. Update command unit test
6. Add a util unit test.
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
fix: #7196
1. Refactor exp_ref_by_name in `dvc/repo/experiments/utils.py`.
In previous the exp_ref_by_name walk throught all of the references to
match only one exp name. This makes multi exp name seeking to be a job
of `O(mn)` complexity, Walking though the exp ref and return all of the
exp ref with greatly reduce this complexity to `O(n)`
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
The function `_raise_error_if_all_disabled` is copyed wrongly to this
place during sub command refactoring. It used to be used in `exp gc`,
can be deleted from this function.
related to #5489
Current dvc/command/experiments.py is too long for switch between cmd class
and the argument parser.
Split the sub command in dvc/command/experiments.py to individual files
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
We didn't mock `os.kill` in `test_dead_process` and the pid it
kills might coincidentlly be the same with some real process.
1. Add a mocker to mock out `os.kill` in `test_dead_process`.
2. Use reset_mock to replace setting it again.
fix: #7062
1. Add `send_signal`, `kill`, `terminate` to processManager.
2. Add tests for each of them.
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
Because our machine name now is the key value of the machine config
dict. And the machine config value will been overwritten by it when the
config is loaded. We can remove the field `name` from the config.
Borrowed this idea from pandas api. Current dropna methods will drop a
column or row if any data is missing. But it is also quite useful to
drop those columns or rows only if all of its elements are missing. This
is mostly used for pruning a table.
fix: #6940
`--global/system/project/local` help message are in some irrelevant
command `status`, `ssh`, `create`, `destroy`. These commands are more
related to running instances rather than `machine config`, level
options didn't work for them.
1. Remove `--global/system/project/local` options in `status`, `ssh`,
`create`, `destroy`.
2. Add a test for it.
3. Correct a typo error.
fix: #6838
1. make machine status name optional.
2. show all machine status if no name provide.
3. UI changes for this cmd.
4. add a new test for it.
Co-authored-by: Dave Berenbaum <[email protected]>
* Add experiment name check.
1. Add experiment name check (https://git-scm.com/docs/git-check-ref-format)
2. Add duplicate exp name check.
3. Add some unit test for it.
* Ban slash / in dvc exp names
* Use dulwich backend for ref name checking
* Some bug fix
* Update dvc/repo/experiments/__init__.py
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
* Update dvc/repo/experiments/__init__.py
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
* Some review changes
* Make some funtion more reusable.
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
* `dvc machine`: add a new command `rename` to it
fix#6481
1. add new command `dvc machine rename`
2. add a unit test for the call.
3. add several functional tests for it.t push myself fix6481
* Move rename logic to tpi
* use tpi.state_mv instead of rename
* Rebase to a newer version
* Update tpi version
* Remove unused assert and try catch
Gao
