Paniagua, Eric wrote:
> Hi Nate,
>
> Thank you for your response! I am glad that it was you in particular who did
> respond, because I also have some questions about the way the upload tool
> handles compressed files and saw that you have opened several Issues related
> to this on the Galaxy bitbucket site. First though, I'll fill you in on my
> further progress on the composite file issue.
>
> As I mentioned in my original email, the trouble is that JobWrapper.finish()
> calls dataset.set_meta() before it calls collect_associated_files(),
> resulting dataset.extra_files_path being inaccurate because the files haven't
> been moved yet from the job working directory. This is all with
> set_metadata_externally=False. (I haven't worked with setting metadata
> externally yet, but I think it is worth verifying whether everything works
> correctly for the case I pointed out when set_metadata_externally=True.)
>
> Since my last email, I poked around a bit more and found that my suggested
> short patch was not correct but incomplete. The core problem is that
> component files are not moved with the primary file, so I changed that (patch
> attached, relative to { https://bitbucket.org/galaxy/galaxy-dist
> 5955:949e4f5fa03a }. Early in JobWrapper.finish() the primary file is moved
> from the working directory to the appropriate directory under
> config.file_path. This patch uses the structure of the path naming
> convention to build the accurate path to the component files, and then moves
> them along with the primary file. It's the least invasive (in terms of
> modifying Galaxy core code) potential fix I came up with, but since it relies
> explicitly on path structure and naming conventions I still think it's a bit
> of a hack. However, it does seem to work, at least to me. I don't have the
> resources or experience yet to attempt to test whether it negatively (or
> positively!) impacts others' use cases!
. I do hope that someone more intimately familiar with the project would be
able to vet it and make improvements / provide feedback / incorporate a fix for
this issue to the Galaxy mainline.
>
> However, since the group I'm working for operates downstream of the main
> Galaxy Development Team (primarily to add tools and maintain a local Galaxy
> server for our institution), we really try to avoid complications in merges
> when pulling down updates by avoiding changing Galaxy core code as much as
> possible. Consequently, I had to shift focus from finding the best correct
> fix to simply finding a workaround, and I found a very simple one. However,
> it relies on the deprecated <code> tool config tag, so it will only work
> until support for the <code> tag is removed. It's self-explanatory and also
> attached.
Hi Eric,
I haven't tried out the patch, but would this have an impact on
processing that occurs when collect_associated_files() runs since they
have been moved back already by your code?
I'd suggest enabling set_metadata_externally = True as I suspect this
will solve the problems since metadata will be set while the primary
file is still in the temporary location. set_metadata_externally = True
should really be the default setting and we may remove internal metadata
setting entirely at some point.
> Regarding potential fixes: Is there any reason why component files of a
> composite dataset should not always follow the primary file? I don't know of
> one, but maybe there is some case in running Galaxy over a cluster where
> there is reason to do otherwise. If the answer is no, it seems the best
> approach is to add an abstraction layer between the datasets and the file
> system to avoid various paths that might be associated to a dataset from
> falling out of sync. One (obviously motivational) function of this layer
> would be to move or copy datasets on the file system atomically. This seems
> pretty important especially for enabling even more flexible support of
> distributed or clustered installations.
We're going to be working on just such an abstraction layer shortly.
> Regarding the upload tool: I just saw your response to my other email on
> Leandro's thread. Thanks for pointing me to the code responsible. I've
> already forked a version of the upload tool for our Galaxy instance. There
> are a lot of issues with the way the way the default upload tool works, and
> it sounds like you're aware of that and perhaps it's been a headache for you
> too. If it might help make progress more swift in the Galaxy mainline, I'd
> be glad to send you patches for updates and changes I make in the upload
> tool. I do have at least one big question about it though: It currently
> attempts to sniff and uncompress files regardless of whether "Auto-detect"
> was selected for "File format"; is that by design or accident? It seems that
> (particularly the bits try uncompressing files) causes headaches for a number
> of developers/users. I'm likely to remove that behavior from my version of
> the upload tool, but I'll be less inclined to do so if that will make it
> diverge si!
gnificantly from Galaxy's expected design path.
We definitely need to make it easier to upload compressed files that
intentionally stay compressed. This should be the case when setting the
specific file type, though.
> Finally, I noted that the <code> tag is deprecated. I learned this from the
> Tool Config Syntax page on the Wiki
> (http://wiki.g2.bx.psu.edu/Admin/Tools/Tool%20Config%20Syntax#A.3Ccode.3E_tag_set).
> It says there that it and the tool hooks (like exec_after_process) are
> being replaced, but doesn't give many details. Is there anything you can
> tell me about what syntax or mechanism is planned to replace them? The
> ability to generate UI parameters, have the tool page itself interact with
> other data sources and programs, use hooks to control tool execution, and
> generally interleave arbitrary code with tool/job runs is very useful.
> Unfortunately, the tools page appears to be out of date or just not remotely
> comprehensive; I've found many examples of syntax in other tools'
> configurations that are simply not documented on the Tool Config Syntax page
> at all.
It's true that <code> is deprecated, although I can't provide much
useful information beyond that. If you post up what you're using hooks
for one of my colleagues with more familiarity with tool configs can
probably reply with the best way to accomplish what you're doing without
hooks.
--nate
>
> Please do let me know if you're able to doing anything (with or without my
> suggestions) about the issue I ran into with the way Galaxy handles composite
> datasets on the file system. Or anything more generally about composite
> dataset design intentions, upload functionality, and known issues. I'd
> greatly appreciate it. Thanks for answering my email!
>
>
> Thanks,
> Eric
>
>
> PS: I'm also unable to download composite datasets that I successfully
> uploaded; instead I get an error page saying the URL path doesn't exist.
> However, I haven't dug into this one enough to think it's not a local
> configuration problem on my end. Do you recognize this as (related to) a
> known issue?
>
> ________________________________________
> From: Nate Coraor [n...@bx.psu.edu]
> Sent: Wednesday, September 28, 2011 9:24 AM
> To: Paniagua, Eric
> Cc: galaxy-dev@lists.bx.psu.edu
> Subject: Re: [galaxy-dev] (Composite) Dataset Upload not Setting Metadata
>
> Paniagua, Eric wrote:
> > Hi all,
> >
> > Can anyone tell me why JobWrapper.finish() moves the primary dataset file
> > dataset_path.false_path to dataset_path.real_path (contingent on
> > config.outputs_to_working_directory == True) but does not move the "extra
> > files"? (lib/galaxy/jobs/__init__.py:540-553) It seems to me that if you
> > want to move a dataset, you want to move the whole dataset, and that
> > perhaps this should be factored out, perhaps into the galaxy.util module?
> >
> > Why does class DatasetPath only account for the path to the primary file
> > and not the path to the "extra files"? It could be used to account for the
> > "extra files" by path splitting as in my previous suggested bug fix, but
> > only if that fix is correct. It doesn't seem to be used for that purpose
> > in the Galaxy code.
> >
> > I look forward to an informative response.
>
> Hi Eric,
>
> Sorry for the delay in responding, and thanks for your very detailed
> digging into this problem. To answer your above question, DatasetPath
> only deals with the primary dataset because extra files are already
> written to the working directory and moved back by the wrapper with
> collect_associated_files().
>
> Making it possible to read these extra files from the working directory
> will be necessary when set_metadata_externally = True in the config,
> although I am surprised this has been broken the whole time. Have you
> made progress past your last email? I can pick up from wherever you've
> left off.
>
> >
> > Thanks,
> > Eric Paniagua
> >
> > ________________________________________
> > From: galaxy-dev-boun...@lists.bx.psu.edu
> > [galaxy-dev-boun...@lists.bx.psu.edu] on behalf of Paniagua, Eric
> > [epani...@cshl.edu]
> > Sent: Monday, September 12, 2011 7:37 PM
> > To: galaxy-dev@lists.bx.psu.edu
> > Subject: Re: [galaxy-dev] (Composite) Dataset Upload not Setting Metadata
> >
> > Hello again,
> >
> > It looks like the config.outputs_to_working_directory variable is intended
> > to do something closely related, but setting it to either of True and False
> > does not in fact fix the problem.
> >
> > The output path for files in a composite dataset upload
> > (dataset.files_path) that is used in the tools/data_source/upload.xml tool
> > is set to a path under the job working directory by
> > lib/galaxy/tools/__init__.py:1519. The preceding code (lines 1507-1516)
> > select the path for the primary file contingent on
> > config.outputs_to_working_directory.
> >
> > Why is the path set in line 1519 not contingent on
> > config.outputs_to_working_directory? Indeed, the following small change
> > fixes the bug I'm observing:
> >
> > diff -r 949e4f5fa03a lib/galaxy/tools/__init__.py
> > --- a/lib/galaxy/tools/__init__.py Mon Aug 29 14:42:04 2011 -0400
> > +++ b/lib/galaxy/tools/__init__.py Mon Sep 12 19:32:26 2011 -0400
> > @@ -1516,7 +1516,9 @@
> > param_dict[name] = DatasetFilenameWrapper( hda )
> > # Provide access to a path to store additional files
> > # TODO: path munging for cluster/dataset server relocatability
> > - param_dict[name].files_path = os.path.abspath(os.path.join(
> > job_working_directory, "dataset_%s_files" % (hda.dataset.id) ))
> > + #param_dict[name].files_path = os.path.abspath(os.path.join(
> > job_working_directory, "dataset_%s_files" % (hda.dataset.id) ))
> > + # This version should make it always follow the primary file
> > + param_dict[name].files_path = os.path.abspath( os.path.join(
> > os.path.split( param_dict[name].file_name )[0], "dataset_%s_files" %
> > (hda.dataset.id) ))
> > for child in hda.children:
> > param_dict[ "_CHILD___%s___%s" % ( name, child.designation
> > ) ] = DatasetFilenameWrapper( child )
> > for out_name, output in self.outputs.iteritems():
> >
> > Would this break anything?
> >
> > If that cannot be changed, would the best solution be to modify the upload
> > tool so that it took care of this on its own? That seems readily doable,
> > but starts to decentralize control of data flow policy.
> >
> > Please advise.
> >
> > Thanks,
> > Eric Paniagua
> > ________________________________________
> > From: galaxy-dev-boun...@lists.bx.psu.edu
> > [galaxy-dev-boun...@lists.bx.psu.edu] on behalf of Paniagua, Eric
> > [epani...@cshl.edu]
> > Sent: Monday, September 12, 2011 1:45 PM
> > To: galaxy-dev@lists.bx.psu.edu
> > Subject: [galaxy-dev] (Composite) Dataset Upload not Setting Metadata
> >
> > Hi everyone,
> >
> > I've been getting my feet wet with Galaxy development working to get some
> > of the rexpression tools online, and I've run into a snag that I've traced
> > back to a set_meta datatype method not being able to find a file from which
> > it wants to extract metadata. After reading the code, I believe this would
> > also be a problem for non-composite datatypes.
> >
> > The specific test case I've been looking at is uploading an affybatch file
> > (and associated pheno file) using Galaxy's built-in upload tool and
> > selecting the File Format manually (ie choosing "affybatch" in the
> > dropdown). I am using unmodified datatype definitions provided in
> > lib/galaxy/datatypes/genetics.py and unmodified core Galaxy upload code as
> > of 5955:949e4f5fa03a. (I am also testing with modified versions, but I am
> > able to reproduce and track this bug in the specified clean version).
> >
> > The crux of the cause of error is that in JobWrapper.finish(),
> > dataset.set_meta() is called (lib/galaxy/jobs/__init__.py:607) before the
> > composite dataset uploaded files are moved (in a call to a Tool method
> > "self.tool.collect_associated_files(out_data, self.working_directory)" on
> > line 670) from the job working directory to the final destination under
> > config.file_path (which defaults to "database/files").
> >
> > In my test case, "database.set_meta( overwrite = False )" eventually calls
> > lib/galaxy/datatypes/genetics.py:Rexp.set_meta(dataset, **kwd). As far as
> > I can tell, the only ways to construct a path to a file (or the file) in a
> > dataset without using hard-coded paths from external knowledge is to use
> > the Dataset.get_file_name or Dataset.extra_files_path properties. Unless
> > explicitly told otherwise, both of these methods construct a path based on
> > the Dataset.file_path class data member, whose value is set during Galaxy
> > startup to config.file_path (default "database/files"). However, at the
> > time set_meta is called in this case, the files are not under
> > config.file_path, but rather under the job working directory. Attempting
> > to open files from the dataset therefore fails when using these paths.
> > However, unless the job working directory is passed to set_meta or during
> > construction of the underlying Dataset object, there doesn't appear to be a
> > way for a Dataset method to acces!
s th!
> > e currently running job (for instance to get its job ID or working
> > directory). (The second suggestion is actually not possible; since the
> > standard upload is asynchronous, the Dataset object is created (and
> > persisted) before the Job that will process it is created.)
> >
> > Thoughts? This issue affects Rexp.set_peek also, as well as any other
> > functions that may want to read data from the uploaded files before they
> > are moved to permanent location. This is why if you have an affybatch file
> > and its associated pheno file and you test this on, say, the public Galaxy
> > server at http://main.g2.bx.psu.edu/ you'll see that the peek info says
> > (for example): "##failed to find
> > /galaxy/main_database/files/002/948/dataset_2948818_files/affybatch_test.pheno"
> >
> > It seems that if the current way that Dataset.file_path, Dataset.file_name,
> > and Dataset.extra_files_path is part of the desired design of Galaxy, that
> > methods like set_meta should be run after the files have been moved to
> > config.file_path so they can set metadata based on file contents. It looks
> > like this is intended to happen at least in some cases, from looking at
> > lib/galaxy/jobs/__init__.py:568-586. However, in my tests this code is not
> > kicking in because hda_tool_output is None.
> >
> > Any clarification on what's happening here, what's supposed to be happening
> > for setting metadata on (potentially composite) uploads, why
> > dataset.set_meta() isn't already being called after the files are moved to
> > config.file_path, or any insights on related Galaxy design decisions I may
> > not know about or design constraints I may have missed would be very
> > greatly appreciated.
> >
> > I'd also be glad to provide further detail or test files upon request.
> >
> > Thank you,
> > Eric Paniagua
> >
> > PS: Further notes on passing the job working directory to set_meta or
> > set_peek - I have been successful modifying the code to do this for
> > set_meta since the call chain starting from dataset.set_meta() in
> > JobWrapper.finish() to Rexp.set_meta() accepts and forwards keyword
> > argument dictionaries along the way. However, set_peek does not accept
> > arbitrary keyword arguments, making it harder to pass along the job working
> > directory when needed without stepping on the toes of any other code.
> >
> > ___________________________________________________________
> > Please keep all replies on the list by using "reply all"
> > in your mail client. To manage your subscriptions to this
> > and other Galaxy lists, please use the interface at:
> >
> > http://lists.bx.psu.edu/
> >
> > ___________________________________________________________
> > Please keep all replies on the list by using "reply all"
> > in your mail client. To manage your subscriptions to this
> > and other Galaxy lists, please use the interface at:
> >
> > http://lists.bx.psu.edu/
> >
> > ___________________________________________________________
> > Please keep all replies on the list by using "reply all"
> > in your mail client. To manage your subscriptions to this
> > and other Galaxy lists, please use the interface at:
> >
> > http://lists.bx.psu.edu/
> # This file provides a workaround to the problem of premature set_meta calls
> in
> # JobWrapper.finish(), which calls set_meta before the component files of a
> # a composite dataset have been moved to their correct paths. This workaround
> # exploits the fact that the exec_after_process hook is called after the
> component
> # files have been moved.
> #
> # This problem affects tools that produce composite datasets in their output.
> #
> # To use this workaround, add the following to your tool's XML configuration
> # (assuming it's in the same directory as this file):
> #
> # <code file="composite_metadata_workaround.py"/>
> #
> # Also, composite types should set peek to start with '##failed' to be
> compatible.
>
> # Retry set_peek and set_meta after the tool has collected all files
> def exec_after_process( app, inp_data, out_data, param_dict,tool, stdout,
> stderr):
> for name, data in out_data.items():
> if data.peek.find( '##failed' ) == 0:
> data.set_meta( overwrite=False )
> data.set_peek( is_multi_byte=data.is_multi_byte() )
> app.model.context.add( data )
> app.model.context.flush()
> diff -r -c -x '.hg*' psu_dist/lib/galaxy/jobs/__init__.py
> glxtemp/lib/galaxy/jobs/__init__.py
> *** psu_dist/lib/galaxy/jobs/__init__.py Fri Sep 16 11:03:48 2011
> --- glxtemp/lib/galaxy/jobs/__init__.py Wed Sep 28 13:58:28 2011
> ***************
> *** 551,556 ****
> --- 551,574 ----
> else:
> self.fail( "Job %s's output dataset(s) could not be
> read" % job.id )
> return
> + efp_fake = dataset_path.false_path.replace( '.dat',
> '_files' )
> + efp_real = dataset_path.real_path.replace( '.dat', '_files'
> )
> + try:
> + shutil.move( efp_fake, efp_real )
> + log.debug( "finish(): Moved %s to %s" % (efp_fake,
> efp_real) )
> + except ( IOError, OSError ) as e:
> + # this can happen if Galaxy is restarted during the
> job's
> + # finish method - the false_path file has already moved,
> + # and when the job is recovered, it won't be found.
> + if os.path.exists( efp_fake ):
> + log.warning( "finish(): Path %s does exist, but
> could not be moved: " % (efp_fake, e.strerror) )
> + else:
> + log.warning( "finish(): Path %s doesn't exist" )
> + if os.path.exists( efp_real ) and os.path.isdir(
> efp_real ):
> + log.warning( "finish(): %s found, so it will be
> used instead" % ( efp_fake, efp_real ) )
> + else:
> + self.fail( "Job %s's output dataset(s) could not be
> read" % job.id )
> + return
> job_context = ExpressionContext( dict( stdout = stdout, stderr =
> stderr ) )
> job_tool = self.app.toolbox.tools_by_id.get( job.tool_id, None )
> def in_directory( file, directory ):
> diff -r -c -x '.hg*' psu_dist/lib/galaxy/tools/__init__.py
> glxtemp/lib/galaxy/tools/__init__.py
> *** psu_dist/lib/galaxy/tools/__init__.py Fri Sep 16 11:03:48 2011
> --- glxtemp/lib/galaxy/tools/__init__.py Wed Sep 28 13:58:44 2011
> ***************
> *** 1516,1522 ****
> param_dict[name] = DatasetFilenameWrapper( hda )
> # Provide access to a path to store additional files
> # TODO: path munging for cluster/dataset server relocatability
> ! param_dict[name].files_path = os.path.abspath(os.path.join(
> job_working_directory, "dataset_%s_files" % (hda.dataset.id) ))
> for child in hda.children:
> param_dict[ "_CHILD___%s___%s" % ( name, child.designation
> ) ] = DatasetFilenameWrapper( child )
> for out_name, output in self.outputs.iteritems():
> --- 1516,1524 ----
> param_dict[name] = DatasetFilenameWrapper( hda )
> # Provide access to a path to store additional files
> # TODO: path munging for cluster/dataset server relocatability
> ! #param_dict[name].files_path = os.path.abspath(os.path.join(
> job_working_directory, "dataset_%s_files" % (hda.dataset.id) ))
> ! # This version should make it always follow the primary file
> ! param_dict[name].files_path = os.path.abspath( os.path.join(
> os.path.split( param_dict[name].file_name )[0], os.path.split(
> param_dict[name].file_name )[1].replace( '.dat', '_files' ) ))
> for child in hda.children:
> param_dict[ "_CHILD___%s___%s" % ( name, child.designation
> ) ] = DatasetFilenameWrapper( child )
> for out_name, output in self.outputs.iteritems():
___________________________________________________________
Please keep all replies on the list by using "reply all"
in your mail client. To manage your subscriptions to this
and other Galaxy lists, please use the interface at:
http://lists.bx.psu.edu/
