pass more information on a dataset merge
Dear list, I thought I was working with fairly large datasets, but they have recently started to include ~2Gb files in sets of >50. I have ran these sort of things before as merged data by using tar to roll them up in one set, but when dealing with >100Gb tarfiles, Galaxy on EC2 seems to get very slow, although that's probably because of my implementation of dataset type detection (untar and read through files). Since tarring/untarring isn't very clean, I want to switch from tarring to creating composite files on merge by putting a tool's results into the dataset.extra_files_path. This doesn't seem to be supported yet, because we currently pass in do_merge the output dataset.filename to the respective datatype's merge method. I would like to pass more data to the merge method (let's say the whole dataset object) to be able to get the composite files directory and 'merge' the files in there. Good idea, bad idea? If anyone has views on this, I'd love to hear them. cheers, jorrit
Jim Johnson and I have been discussing that approach to handling fractionated proteomics samples as well (composite datatypes, not the specifics of the interface for parallelizing). My perspective has been that Galaxy should be augmented with better native mechanisms for grouping objects in histories, operating over those groups, building workflows that involve arbitrary numbers of inputs, etc... Composite data types are kindof a kludge, I think they are more useful for grouping HTML files together when you don't care about operating on the constituent parts you just want to view pages a as a report or something. With this proteomic data we are working with, the individual pieces are really interesting right? You want to operate on the individual pieces with the full array of tools (not just these special tools that have the logic for dealing with the composite datatypes), you want to visualize the files, etc... Putting these component pieces in the composite data type extra_files path really limits what you can do with the pieces in Galaxy. I have a vague idea of something that I think could bridge some of the gaps between the approaches (though I have no clue on the feasibility). Looking through your implementation on bitbucket it looks like you are defining your core datatypes (MS2, CruxSequest) as subclasses of this composite data type (CompositeMultifile). My recommendation would be to try to define plain datatypes for these core datatype (MS2, CruxSequest) and then have the separate composite datatype sort of delegate to the plain datatypes. You could then continue to explicitly declare subclasses of the composite datatype (maybe MS2Set, CruxSequestSet), but also maybe augement the tool xml so you can do implicit data type instances the way you can with tabular data for instance (instead of defining columns you would define the datatype to delegate to). The next step would be to make the parallelism implicit (i.e pull it out of the tool wrapper). Your tool wrappers wouldn't reference the composite datatypes, they would reference the simple datatypes, but you could add a little icon next to any input that let you replace a single input with a composite input for that type. It would be kind of like the run workflow page where you can replace an input with a multiple inputs. If a composite input (or inputs) are selected the tool would then produce composite outputs. For the steps that actually combine multiple inputs, I think in your case this is perculator maybe (a tool like interprophet or Scaffold that merges peptide probabilities across runs and groups proteins), then you could have the same sort of implicit replacement but instead of for single inputs it could do that for multi-inputs (assuming the Galaxy powers that be accept my fixes for multi-input tool parameters: https://bitbucket.org/galaxy/galaxy-central/pull-request/76/multi-input-data...). The upshot of all of that would be that then even if these composites datatypes aren't used widely, other people could still use your proteomics tools (my users are definitely interested in Crux for instance) and you could then use other developers' proteomic tools with your composite datatypes even though they weren't designed with that use case in mind (I have msconvert, myrimatch, idpicker, proteinpilot, Ira Cooke has X! Tandem, OMSSA, TPP, and NBIC has an entire suite of label free quant tools). A third benefit would be that people working in other -omicses could make use of the homogenous composite datatype implementation without needing to rewrite their wrappers and datatypes. There is probably something that I am missing that makes this very difficult, let me know if you think this is a good idea and what its feasibility might be. I forked your repo and set off to try to implement some of this stuff last week and I ended up with my galaxy pull requests to improve batching workflows and multi-input tool parameters instead, but I hope to eventually get around to it. -John ------------------------------------------------ John Chilton Senior Software Developer University of Minnesota Supercomputing Institute Office: 612-625-0917 Cell: 612-226-9223 Bitbucket: https://bitbucket.org/jmchilton Github: https://github.com/jmchilton Web: http://jmchilton.net On Mon, Oct 1, 2012 at 8:24 AM, Jorrit Boekel <jorrit.boekel@scilifelab.se> wrote:
Dear list,
I thought I was working with fairly large datasets, but they have recently started to include ~2Gb files in sets of >50. I have ran these sort of things before as merged data by using tar to roll them up in one set, but when dealing with >100Gb tarfiles, Galaxy on EC2 seems to get very slow, although that's probably because of my implementation of dataset type detection (untar and read through files).
Since tarring/untarring isn't very clean, I want to switch from tarring to creating composite files on merge by putting a tool's results into the dataset.extra_files_path. This doesn't seem to be supported yet, because we currently pass in do_merge the output dataset.filename to the respective datatype's merge method.
I would like to pass more data to the merge method (let's say the whole dataset object) to be able to get the composite files directory and 'merge' the files in there. Good idea, bad idea? If anyone has views on this, I'd love to hear them.
cheers, jorrit
___________________________________________________________ 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:
Here is an implementation of the implicit multi-file composite datatypes piece of that idea. I think the implicit parallelism may be harder. https://bitbucket.org/galaxyp/galaxy-central-homogeneous-composite-datatypes... Jorrit do you have any objection to me trying to get this included in galaxy-central (this is 95% code I stole from you)? I made the changes against a clean galaxy-central fork and included nothing proteomics specific in anticipation of trying to do that. I have talked with Jim Johnson about the idea and he believes it would be useful his mothur metagenomics tools, so the idea is valuable outside of proteomics. Galaxy team, would you be okay with including this and if so is there anything you would like to see either at a high level or at the level of the actual implementation. -John ------------------------------------------------ John Chilton Senior Software Developer University of Minnesota Supercomputing Institute Office: 612-625-0917 Cell: 612-226-9223 Bitbucket: https://bitbucket.org/jmchilton Github: https://github.com/jmchilton Web: http://jmchilton.net On Mon, Oct 8, 2012 at 9:24 AM, John Chilton <chilton@msi.umn.edu> wrote:
Jim Johnson and I have been discussing that approach to handling fractionated proteomics samples as well (composite datatypes, not the specifics of the interface for parallelizing).
My perspective has been that Galaxy should be augmented with better native mechanisms for grouping objects in histories, operating over those groups, building workflows that involve arbitrary numbers of inputs, etc... Composite data types are kindof a kludge, I think they are more useful for grouping HTML files together when you don't care about operating on the constituent parts you just want to view pages a as a report or something. With this proteomic data we are working with, the individual pieces are really interesting right? You want to operate on the individual pieces with the full array of tools (not just these special tools that have the logic for dealing with the composite datatypes), you want to visualize the files, etc... Putting these component pieces in the composite data type extra_files path really limits what you can do with the pieces in Galaxy.
I have a vague idea of something that I think could bridge some of the gaps between the approaches (though I have no clue on the feasibility). Looking through your implementation on bitbucket it looks like you are defining your core datatypes (MS2, CruxSequest) as subclasses of this composite data type (CompositeMultifile). My recommendation would be to try to define plain datatypes for these core datatype (MS2, CruxSequest) and then have the separate composite datatype sort of delegate to the plain datatypes.
You could then continue to explicitly declare subclasses of the composite datatype (maybe MS2Set, CruxSequestSet), but also maybe augement the tool xml so you can do implicit data type instances the way you can with tabular data for instance (instead of defining columns you would define the datatype to delegate to).
The next step would be to make the parallelism implicit (i.e pull it out of the tool wrapper). Your tool wrappers wouldn't reference the composite datatypes, they would reference the simple datatypes, but you could add a little icon next to any input that let you replace a single input with a composite input for that type. It would be kind of like the run workflow page where you can replace an input with a multiple inputs. If a composite input (or inputs) are selected the tool would then produce composite outputs.
For the steps that actually combine multiple inputs, I think in your case this is perculator maybe (a tool like interprophet or Scaffold that merges peptide probabilities across runs and groups proteins), then you could have the same sort of implicit replacement but instead of for single inputs it could do that for multi-inputs (assuming the Galaxy powers that be accept my fixes for multi-input tool parameters: https://bitbucket.org/galaxy/galaxy-central/pull-request/76/multi-input-data...).
The upshot of all of that would be that then even if these composites datatypes aren't used widely, other people could still use your proteomics tools (my users are definitely interested in Crux for instance) and you could then use other developers' proteomic tools with your composite datatypes even though they weren't designed with that use case in mind (I have msconvert, myrimatch, idpicker, proteinpilot, Ira Cooke has X! Tandem, OMSSA, TPP, and NBIC has an entire suite of label free quant tools). A third benefit would be that people working in other -omicses could make use of the homogenous composite datatype implementation without needing to rewrite their wrappers and datatypes.
There is probably something that I am missing that makes this very difficult, let me know if you think this is a good idea and what its feasibility might be. I forked your repo and set off to try to implement some of this stuff last week and I ended up with my galaxy pull requests to improve batching workflows and multi-input tool parameters instead, but I hope to eventually get around to it.
-John
------------------------------------------------ John Chilton Senior Software Developer University of Minnesota Supercomputing Institute Office: 612-625-0917 Cell: 612-226-9223 Bitbucket: https://bitbucket.org/jmchilton Github: https://github.com/jmchilton Web: http://jmchilton.net
On Mon, Oct 1, 2012 at 8:24 AM, Jorrit Boekel <jorrit.boekel@scilifelab.se> wrote:
Dear list,
I thought I was working with fairly large datasets, but they have recently started to include ~2Gb files in sets of >50. I have ran these sort of things before as merged data by using tar to roll them up in one set, but when dealing with >100Gb tarfiles, Galaxy on EC2 seems to get very slow, although that's probably because of my implementation of dataset type detection (untar and read through files).
Since tarring/untarring isn't very clean, I want to switch from tarring to creating composite files on merge by putting a tool's results into the dataset.extra_files_path. This doesn't seem to be supported yet, because we currently pass in do_merge the output dataset.filename to the respective datatype's merge method.
I would like to pass more data to the merge method (let's say the whole dataset object) to be able to get the composite files directory and 'merge' the files in there. Good idea, bad idea? If anyone has views on this, I'd love to hear them.
cheers, jorrit
___________________________________________________________ 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:
Hi John, I tried your galaxy-central-homogeneous-composite-datatypes implementation, works great thank you (and Jorrit). A couple of fixes: 1. Add multi_upload.xml to too_conf.xml 2. lib/galaxy/tools/parameters/grouping.py line 322 (in get_filenames( context )) - "if ftp_files is not None:" Remove "is not None" as ftp_files is empty [], but not None, then line 331 "user_ftp_dir = os.path.join( trans.app.config.ftp_upload_dir, trans.user.email )" throws an exeption if ftp_upload_dir isn't set. Alex -----Original Message----- From: galaxy-dev-bounces@lists.bx.psu.edu [mailto:galaxy-dev-bounces@lists.bx.psu.edu] On Behalf Of John Chilton Sent: Tuesday, 16 October 2012 1:07 AM To: Jorrit Boekel Cc: galaxy-dev@lists.bx.psu.edu Subject: Re: [galaxy-dev] pass more information on a dataset merge Here is an implementation of the implicit multi-file composite datatypes piece of that idea. I think the implicit parallelism may be harder. https://bitbucket.org/galaxyp/galaxy-central-homogeneous-composite-datatypes... Jorrit do you have any objection to me trying to get this included in galaxy-central (this is 95% code I stole from you)? I made the changes against a clean galaxy-central fork and included nothing proteomics specific in anticipation of trying to do that. I have talked with Jim Johnson about the idea and he believes it would be useful his mothur metagenomics tools, so the idea is valuable outside of proteomics. Galaxy team, would you be okay with including this and if so is there anything you would like to see either at a high level or at the level of the actual implementation. -John ------------------------------------------------ John Chilton Senior Software Developer University of Minnesota Supercomputing Institute Office: 612-625-0917 Cell: 612-226-9223 Bitbucket: https://bitbucket.org/jmchilton Github: https://github.com/jmchilton Web: http://jmchilton.net On Mon, Oct 8, 2012 at 9:24 AM, John Chilton <chilton@msi.umn.edu> wrote:
Jim Johnson and I have been discussing that approach to handling fractionated proteomics samples as well (composite datatypes, not the specifics of the interface for parallelizing).
My perspective has been that Galaxy should be augmented with better native mechanisms for grouping objects in histories, operating over those groups, building workflows that involve arbitrary numbers of inputs, etc... Composite data types are kindof a kludge, I think they are more useful for grouping HTML files together when you don't care about operating on the constituent parts you just want to view pages a as a report or something. With this proteomic data we are working with, the individual pieces are really interesting right? You want to operate on the individual pieces with the full array of tools (not just these special tools that have the logic for dealing with the composite datatypes), you want to visualize the files, etc... Putting these component pieces in the composite data type extra_files path really limits what you can do with the pieces in Galaxy.
I have a vague idea of something that I think could bridge some of the gaps between the approaches (though I have no clue on the feasibility). Looking through your implementation on bitbucket it looks like you are defining your core datatypes (MS2, CruxSequest) as subclasses of this composite data type (CompositeMultifile). My recommendation would be to try to define plain datatypes for these core datatype (MS2, CruxSequest) and then have the separate composite datatype sort of delegate to the plain datatypes.
You could then continue to explicitly declare subclasses of the composite datatype (maybe MS2Set, CruxSequestSet), but also maybe augement the tool xml so you can do implicit data type instances the way you can with tabular data for instance (instead of defining columns you would define the datatype to delegate to).
The next step would be to make the parallelism implicit (i.e pull it out of the tool wrapper). Your tool wrappers wouldn't reference the composite datatypes, they would reference the simple datatypes, but you could add a little icon next to any input that let you replace a single input with a composite input for that type. It would be kind of like the run workflow page where you can replace an input with a multiple inputs. If a composite input (or inputs) are selected the tool would then produce composite outputs.
For the steps that actually combine multiple inputs, I think in your case this is perculator maybe (a tool like interprophet or Scaffold that merges peptide probabilities across runs and groups proteins), then you could have the same sort of implicit replacement but instead of for single inputs it could do that for multi-inputs (assuming the Galaxy powers that be accept my fixes for multi-input tool parameters: https://bitbucket.org/galaxy/galaxy-central/pull-request/76/multi-input-data...).
The upshot of all of that would be that then even if these composites datatypes aren't used widely, other people could still use your proteomics tools (my users are definitely interested in Crux for instance) and you could then use other developers' proteomic tools with your composite datatypes even though they weren't designed with that use case in mind (I have msconvert, myrimatch, idpicker, proteinpilot, Ira Cooke has X! Tandem, OMSSA, TPP, and NBIC has an entire suite of label free quant tools). A third benefit would be that people working in other -omicses could make use of the homogenous composite datatype implementation without needing to rewrite their wrappers and datatypes.
There is probably something that I am missing that makes this very difficult, let me know if you think this is a good idea and what its feasibility might be. I forked your repo and set off to try to implement some of this stuff last week and I ended up with my galaxy pull requests to improve batching workflows and multi-input tool parameters instead, but I hope to eventually get around to it.
-John
------------------------------------------------ John Chilton Senior Software Developer University of Minnesota Supercomputing Institute Office: 612-625-0917 Cell: 612-226-9223 Bitbucket: https://bitbucket.org/jmchilton Github: https://github.com/jmchilton Web: http://jmchilton.net
On Mon, Oct 1, 2012 at 8:24 AM, Jorrit Boekel <jorrit.boekel@scilifelab.se> wrote:
Dear list,
I thought I was working with fairly large datasets, but they have recently started to include ~2Gb files in sets of >50. I have ran these sort of things before as merged data by using tar to roll them up in one set, but when dealing with >100Gb tarfiles, Galaxy on EC2 seems to get very slow, although that's probably because of my implementation of dataset type detection (untar and read through files).
Since tarring/untarring isn't very clean, I want to switch from tarring to creating composite files on merge by putting a tool's results into the dataset.extra_files_path. This doesn't seem to be supported yet, because we currently pass in do_merge the output dataset.filename to the respective datatype's merge method.
I would like to pass more data to the merge method (let's say the whole dataset object) to be able to get the composite files directory and 'merge' the files in there. Good idea, bad idea? If anyone has views on this, I'd love to hear them.
cheers, jorrit
___________________________________________________________ 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:
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/
No objections whatsoever, I'm really happy more people are interested! General fixes are definitely to be preferred over a whatever-field-specific solution if you ask me. I am currently running a parallelism where I create symbolic links on split, and move result files (as opposed to copying) on a merge. Faster than copying back and forward, but it's limited to splitting to the amount of files in a set. cheers, jorrit On 10/15/2012 04:07 PM, John Chilton wrote:
Here is an implementation of the implicit multi-file composite datatypes piece of that idea. I think the implicit parallelism may be harder.
https://bitbucket.org/galaxyp/galaxy-central-homogeneous-composite-datatypes...
Jorrit do you have any objection to me trying to get this included in galaxy-central (this is 95% code I stole from you)? I made the changes against a clean galaxy-central fork and included nothing proteomics specific in anticipation of trying to do that. I have talked with Jim Johnson about the idea and he believes it would be useful his mothur metagenomics tools, so the idea is valuable outside of proteomics.
Galaxy team, would you be okay with including this and if so is there anything you would like to see either at a high level or at the level of the actual implementation.
-John
------------------------------------------------ John Chilton Senior Software Developer University of Minnesota Supercomputing Institute Office: 612-625-0917 Cell: 612-226-9223 Bitbucket: https://bitbucket.org/jmchilton Github: https://github.com/jmchilton Web: http://jmchilton.net
On Mon, Oct 8, 2012 at 9:24 AM, John Chilton <chilton@msi.umn.edu> wrote:
Jim Johnson and I have been discussing that approach to handling fractionated proteomics samples as well (composite datatypes, not the specifics of the interface for parallelizing).
My perspective has been that Galaxy should be augmented with better native mechanisms for grouping objects in histories, operating over those groups, building workflows that involve arbitrary numbers of inputs, etc... Composite data types are kindof a kludge, I think they are more useful for grouping HTML files together when you don't care about operating on the constituent parts you just want to view pages a as a report or something. With this proteomic data we are working with, the individual pieces are really interesting right? You want to operate on the individual pieces with the full array of tools (not just these special tools that have the logic for dealing with the composite datatypes), you want to visualize the files, etc... Putting these component pieces in the composite data type extra_files path really limits what you can do with the pieces in Galaxy.
I have a vague idea of something that I think could bridge some of the gaps between the approaches (though I have no clue on the feasibility). Looking through your implementation on bitbucket it looks like you are defining your core datatypes (MS2, CruxSequest) as subclasses of this composite data type (CompositeMultifile). My recommendation would be to try to define plain datatypes for these core datatype (MS2, CruxSequest) and then have the separate composite datatype sort of delegate to the plain datatypes.
You could then continue to explicitly declare subclasses of the composite datatype (maybe MS2Set, CruxSequestSet), but also maybe augement the tool xml so you can do implicit data type instances the way you can with tabular data for instance (instead of defining columns you would define the datatype to delegate to).
The next step would be to make the parallelism implicit (i.e pull it out of the tool wrapper). Your tool wrappers wouldn't reference the composite datatypes, they would reference the simple datatypes, but you could add a little icon next to any input that let you replace a single input with a composite input for that type. It would be kind of like the run workflow page where you can replace an input with a multiple inputs. If a composite input (or inputs) are selected the tool would then produce composite outputs.
For the steps that actually combine multiple inputs, I think in your case this is perculator maybe (a tool like interprophet or Scaffold that merges peptide probabilities across runs and groups proteins), then you could have the same sort of implicit replacement but instead of for single inputs it could do that for multi-inputs (assuming the Galaxy powers that be accept my fixes for multi-input tool parameters: https://bitbucket.org/galaxy/galaxy-central/pull-request/76/multi-input-data...).
The upshot of all of that would be that then even if these composites datatypes aren't used widely, other people could still use your proteomics tools (my users are definitely interested in Crux for instance) and you could then use other developers' proteomic tools with your composite datatypes even though they weren't designed with that use case in mind (I have msconvert, myrimatch, idpicker, proteinpilot, Ira Cooke has X! Tandem, OMSSA, TPP, and NBIC has an entire suite of label free quant tools). A third benefit would be that people working in other -omicses could make use of the homogenous composite datatype implementation without needing to rewrite their wrappers and datatypes.
There is probably something that I am missing that makes this very difficult, let me know if you think this is a good idea and what its feasibility might be. I forked your repo and set off to try to implement some of this stuff last week and I ended up with my galaxy pull requests to improve batching workflows and multi-input tool parameters instead, but I hope to eventually get around to it.
-John
------------------------------------------------ John Chilton Senior Software Developer University of Minnesota Supercomputing Institute Office: 612-625-0917 Cell: 612-226-9223 Bitbucket: https://bitbucket.org/jmchilton Github: https://github.com/jmchilton Web: http://jmchilton.net
On Mon, Oct 1, 2012 at 8:24 AM, Jorrit Boekel <jorrit.boekel@scilifelab.se> wrote:
Dear list,
I thought I was working with fairly large datasets, but they have recently started to include ~2Gb files in sets of >50. I have ran these sort of things before as merged data by using tar to roll them up in one set, but when dealing with >100Gb tarfiles, Galaxy on EC2 seems to get very slow, although that's probably because of my implementation of dataset type detection (untar and read through files).
Since tarring/untarring isn't very clean, I want to switch from tarring to creating composite files on merge by putting a tool's results into the dataset.extra_files_path. This doesn't seem to be supported yet, because we currently pass in do_merge the output dataset.filename to the respective datatype's merge method.
I would like to pass more data to the merge method (let's say the whole dataset object) to be able to get the composite files directory and 'merge' the files in there. Good idea, bad idea? If anyone has views on this, I'd love to hear them.
cheers, jorrit
___________________________________________________________ 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:
participants (4)
-
Alex.Khassapov＠csiro.au -
John Chilton -
John Chilton -
Jorrit Boekel