wip! workflows as programs prototyping

[tsibley]

See commit messages. Opening for visibility and soliciting commentary. Not for merge yet.

Checklist

• Checks pass

[jameshadfield]

The behaviour of multiple configfiles is somewhat documented but I think will prove a tripping hazard to people.

Multiple configfiles (as per this section of code, and/or via additional --configfiles) will overwrite scalars and lists; dicts will be merged, but nested dicts won't be; there's no way to remove a dictionary key as far as I can find.

This has implications for how we structure configs, specifically around how we encode flags to skip rules or skip a particular argument.

[tsibley]

Nod, for sure, but "multiple config files" is already our established approach, so this is making that more accessible/standard but isn't changing approaches.

We could change approaches, though, by combining the configs ourselves outside of Snakemake's machinery. I didn't think there was appetite for that, but it's very possible.

dicts will be merged, but nested dicts won't be; there's no way to remove a dictionary key as far as I can find.

Hmm. In my recollection (and testing right now), nested dicts are merged? That is, the config merge is deep.

The merge is additive though, so yes, there is no way to remove keys.

[jameshadfield]

nested dicts are merged

Oh yeah, I wasn't formatting my YAML correctly. That they are merged will be very helpful for us.

[jameshadfield]

"multiple config files" is already our established approach, so this is making that more accessible/standard but isn't changing approaches.

I can see the point of view that this isn't a workflows-as-programs thing rather a pathgon-repo-{guide,template} thing. Similar to my comment in a recent meeting about merging in additional data I think these concepts should be thought of as intertwined rather than isolated in order to make workflows-as-programs as successful as possible.

Our current use of multiple config files are largely just adding in specific steps, e.g I think the examples in measles are a good summary of how we use them more generally:

• Measles automated builds use an additional config to add in a deploy rule and the associated config params
• Measles CI uses an additional config to add a rule which copies example data and thus avoids downloading the entire dataset.

I think it's rarer¹ to use additional configs to change parameters of builds, or to subset build targets etc. If we want to recommend this approach -- an approach that I think we should use and workflows-as-programs seems to lead us towards anyway -- then it's not going to play nicely with our current config structures. Some examples:

• Avian-flu uses a~nested-dict to define the builds to run (subtype x time window). So we can't override this with a config file without having some approach to encoding "skip this" / "no-op". (Mentioned here as well.) Edit: actually in its current formation it's a dict of lists so there is a way to overwrite, albeit a little clunky.
• Seasonal flu (at least, the public builds) uses an array of segments and a dictionary of builds (builds here is h1n1pdm_2y, h3n2_2y etc). So config overlays aren't going to be able to target a subset of those builds.
• RSV uses two lists to encode the builds & resolutions which will play nicely with config overrides but is predicated on all builds being run for all resolutions which isn't always the case.
• (Somewhat unrelated, but...) Dengue hardcodes serotypes/genes so that'll need lifting into config to make the most of workflows-as-programs.

cc @joverlee521

¹ The most complex workflows are seasonal-flu and ncov, so maybe those utilise the style of config merging I'm talking about here. I don't think seasonal-flu does at all, but maybe? It'd be great to have examples in Nextstrain where we do make use of config merging to replace / subset config parameters.

[genehack]

"multiple config files" is already our established approach, so this is making that more accessible/standard but isn't changing approaches.

...

I think it's rarer¹ to use additional configs to change parameters of builds, or to subset build targets etc. If we want to recommend this approach -- an approach that I think we should use and workflows-as-programs seems to lead us towards anyway -- then it's not going to play nicely with our current config structures. Some examples:

(avain-flu, seasonal-flu, RSV examples elided)

I think it's totally fair to draw a line, somewhere (not exactly sure where), and say the pathogen repos on one side of it are "legacy" and will continue to be maintained the way they have been historically, and will not use these new patterns, while pathogen repos on the other side of the line are "current" and will be keep reasonably up-to-date and consistent with our best practices. (We can also decide, down the road, to invest the effort into converting "legacy" repos into "current" repos, if and when that feels like it has a positive return-on-effort.)

Note: not currently taking a strong stance on where that line is 😁

* (Somewhat unrelated, but...) Dengue [hardcodes](https://github.com/nextstrain/dengue/blob/6f9144f7767c0cac4755c60cd5c069a5d2c4ed6f/phylogenetic/Snakefile#L3-L4) serotypes/genes so that'll need lifting into config to make the most of workflows-as-programs.

Note that measles and yellow fever do the same thing.

[jameshadfield]

An example of this is the WNV/washington-state config, which layers on top of the default config.

Thanks for the example! That's a great example of a simple config structure (i.e. a single dataset target, no wildcards in the workflow) where merging works nicely. It will be interesting to see how that develops as WNV expands (e.g. to have NA & global targets) - does it go the mpox route of one config (or one config overlay) per target, or does it tend towards a more complicated config structure. It's specifically the more complicated structures I was focusing on above.

(It also uses an approach I think we've moved away from, namely

refine:
  treetime_params: --coalescent opt --date-inference marginal ...

which makes it straightforward to remove/add options/params as you're in charge of the entire string. As a thought exercise, if the base config instead specified

refine:
  coalescent: opt

and the overlay config wanted to skip that argument entirely how would it do it?)

[tsibley]

As a thought exercise, if the base config instead specified

refine:
  coalescent: opt

and the overlay config wanted to skip that argument entirely how would it do it?)

I can't for the life of me find where I wrote about this recentlyish, but I'd suggested previously:

refine:
  coalescent: ~   # or: null

That's a great example of a simple config structure (i.e. a single dataset target, no wildcards in the workflow) where merging works nicely. It will be interesting to see how that develops as WNV expands (e.g. to have NA & global targets) - does it go the mpox route of one config (or one config overlay) per target, or does it tend towards a more complicated config structure.

The niceness of simple config structures is why, in Slack recently, I suggested we consider "small multiples" of config (optionally with tooling to allow derived configs/inheritance) rather than a large complex cross-product config.

[tsibley]

I can't for the life of me find…

oh, it was a freakin' Google Docs comment.

[tsibley]

dicts will be merged, but nested dicts won't be

P.S. This behaviour changed in Snakemake 6.11.0 (Nov 2021). Before that nested dicts were not merged. (Yep, it was another potentially-breaking change in a minor version release.)

[jameshadfield]

but I'd suggested previously:

Yeah, saw that. I think something like that's the right move, and the corresponding snakemake logic needs to ensure that it drops the --coalescent argument name as well as the null value. (We can probably use a helpful helper function here.)

I prototyped a similar approach for nextstrain/avian-flu#104 but in the context of defining what builds to run (for gisaid there are 48 by default, so it's nice to (a) not enumerate them all and (b) be able to subset them for a particular analysis). One of the prototypes looked like this:

builds:
  h5nx:
    all-time:
      - ha
      - na
    2y: ∅
  h5n1: ∅
  h7n9: ∅
  h9n2: ∅

But I didn't like the necessity of listing out all the builds you don't want to run. I switched to a list-based interface which is much nicer in this respect. (There were other reasons for switching too, namely the hierarchy here isn't right and to make it right results in very verbose dictionaries.)

[jameshadfield]

I'd like to see config parsing such as this become part of our general snakemake workflows regardless of workflows as programs CLI work.

[tsibley]

I'd like to see config parsing such as this become part of our general snakemake workflows regardless of workflows as programs CLI work.

Agreed! (ever since I started 🙃)

[joverlee521]

My main concern is the traceability of config files with this change. I think I need to wrap my head around the fact that the same exact config file can result in different input files depending on which working directory was used.

[joverlee521]

Slightly concerned that this change may confuse external users who want to trace the default files. It's not immediately clear that these are pulling from "defaults/".

[tsibley]

So the ingest/defaults/config.yaml comments on files' locations, but this phylogenetic/defaults/config.yaml did not and so I also did not bother. But we very easily could if we think it'll be a tripping hazard.

If we're terribly concerned and don't think a comment is enough, we can also leave the defaults/ prefix in the default config. It's handled for backwards-compat.

measles/shared/functions.smk

Lines 20 to 24 in 6584e68

	# Special-case defaults/… for backwards compatibility with older
	# configs. We could achieve the same behaviour with a symlink
	# (defaults/defaults → .) but that seems less clear.
	if path.startswith("defaults/"):
	defaults_path = os.path.join(workflow.basedir, path)

[joverlee521]

I think the additional comments should be enough.

we can also leave the defaults/ prefix in the default config.

I thought the defaults/ prefix was removed so that the config will automatically use input files in the analysis directory? If we keep the defaults/ prefix, users must override the config param or put their files in a defaults/ directory within their analysis directory right?

[jameshadfield]

I thought the defaults/ prefix was removed so that the config will automatically use input files in the analysis directory?

That's my understanding as well.

The search order is:

1. Filename in current analysis directory (if it exists)
2. Filename in workflow's defaults/ directory. (If the config-specified filename has "defaults/" then that's ignored for this step.)

By not using defaults/ in the config-specified filename it makes it possible to swap out the workflow file (TSV etc) file with one of the same name in the current analysis directory.

[tsibley]

I thought the defaults/ prefix was removed so that the config will automatically use input files in the analysis directory?

Sorry, yes, you're both correct with the code in this PR as-is and what I said. The bit I forgot to say as I was rushing out the door (ahh! 🤦) was that we could strip the defaults/ prefix off when searching for files in the analysis directory, thus letting us keep the defaults/ prefix in the default config. Given we're already special-casing it for backwards-compat, we could instead special-case it the other way for analysis directory lookup instead.

(This is actually the first pattern I tinkered with over a month ago, but I thought it was more confusing than a simple unchanging path + search order.)

[tsibley]

There may be more than one, of course.

Hmm? There's only one config object at run time, and that's what we'd write out.

[jameshadfield]

If I run multiple invocations of the workflow with different overlays, or invocations of different workflows, each will write sets of files into the analysis directory, and these sets may not overlap or may partially overlap. To use an avian-flu example, I may have an analysis directory with both gisaid & cattle-outbreak datasets.

[tsibley]

Ah, I see. Multiple invocations (whether same workflow or not) writing to the same working analysis directory seemed to me like something we didn't need to support (in terms of keeping things separate): there's always the chance of overwriting files in that case (often overwriting is desired/expected) and confusion around what's output from which invocation. Do you think we definitely do need to support that sort of use case? If so, the implications of that would seem to extend far beyond the proposed results/config.yaml discussed here.

[jameshadfield]

Do you think we definitely do need to support that sort of use case?

No - but it's something to keep in mind as the structure is built out, because what we're building out does support it and I think it'll be a common use case. For instance, if we could include some identifier based on the invocation within the filename we wouldn't be simply overwriting results/config.yaml each time. This follows on from my comment above, I think we should also write out information beyond the config, e.g. pathogen version used, runtime version, invocation command etc. You could shoehorn that into the written-out config, or create a directory for each workflow invocation with files detailing all that information; I'd think logs & benchmarks may better live here as well. Obviously we don't have to do all this now, but thinking about the structure now seems like the right move.

[tsibley]

For instance, if we could include some identifier based on the invocation within the filename we wouldn't be simply overwriting results/config.yaml each time.

Right, of course, but we're still overwriting all/most of the other files in results/ each time though. I don't see the harm in matching that behaviour with results/config.yaml for now.

I think we should also write out information beyond the config, e.g. pathogen version used, runtime version, invocation command etc. You could shoehorn that into the written-out config, or create a directory for each workflow invocation with files detailing all that information; I'd think logs & benchmarks may better live here as well. Obviously we don't have to do all this now, but thinking about the structure now seems like the right move.

A timestamped directory to contain all this individual-run information makes the most sense to me and is the most conventional… but I don't think we need to worry about this structure right now now. I think it's too early to implement now; we should focus on the core workflows-as-programs changes first before building on them while they're still forming. I also think there's some consideration to take on what outputs get differentiated between invocations (e.g. protected from overwrite) and what outputs don't, and why that split is useful vs. the default situation of "nothing's protected" or "everything's protected".

[joverlee521]

Slightly concerned that users will be confused by the term "working directory", especially those who are not familiar with Snakemake. (I had originally made everything relative to the ingest directory so that users had a concrete reference point.)

[tsibley]

For user reference, we can say "your analysis directory", perhaps with an explanatory parenthetical like "(Snakemake's working directory, --directory, often your current directory)". This seems like a tripping hazard we can address various ways.

[joverlee521]

This seems like a tripping hazard we can address various ways.

Totally! Just want to make sure we address it as paths are often the most confusing thing...

Seems like the working directory can be different things depending on how you run the workflow.

1. native snakemake = the current directory or provided --directory.
2. nextstrain build <directory> = the provided <directory> (can be complicated by additional Snakemake --directory option)
3. nextstrain run <pathogen-name> <workflow-name> <analysis-directory> = provided <analysis-directory>

I do like that (3) is the most straight-forward so all for moving in that direction.

[jameshadfield]

Following this up after work in avian-flu and discussions in slack, I think we'll need words to describe the following 4 locations (even if some are only for developers). This is assuming we go with the slight variant to (3) which allows things like nextstrain run avian-flu phylogenetic/h5n1-cattle-outbreak my-analysis.

1. "pathogen-name" - maybe obvious, but this may include version identifiers etc
2. the workflow name - "ingest", "phylogenetic", "phylogenetic/gisaid", "phylogenetic/clade-i" etc. workflow.basedir in snakemake.
3. the analysis directory. os.getcwd() in snakemake.
4. The "base" workflow directory if there are (sub-)workflows which use a common snakefile. In avian-flu this'd be avian-flu/Snakefile (or maybe one day avian-flu/phylogenetic/Snakefile). This should only be relevant for workflow developers.

[tsibley]

1. The pathogen spec I'm working from in Nextstrain CLI is <name>[@<version>[=<url>]]. A URL is only accepted by nextstrain setup, as a way to override the default assumption that the name refers to a Nextstrain pathogen repo and/or that the version refers to a Git rev in the same.

2. I want to stress that I think it's important to conceptually treat/talk about these mostly as names (like you've done!) and not paths, even if they're used internally as paths initially (and likely indefinitely). Doing so makes it more sensible to do things like alias (or "redirect") old names to new names (e.g. via entries in nextstrain-pathogen.yaml, or symlinks) to ease use or preserve backwards compatibility.

3. I'd call this the "working analysis directory" in full, shortening to "analysis directory" (for Nextstrain CLI) or "working directory" (for Snakemake development).

4. I think avian-flu/Snakefile (or avian-flu/phylogenetic/Snakefile) would ideally be avian-flu/phylogenetic/rules/base.smk (or core.smk or something similar). That is, we'd stick to the convention of only using the Snakefile filename for supported workflow entrypoints. I think this will reduce confusion.

[tsibley]

I think I need to wrap my head around the fact that the same exact config file can result in different input files depending on which working directory was used.

Nod. One upside here is that we can say things like "send us a copy of your analysis directory (working directory)" instead of "send us your config, plus any additional files you configured".

[genehack]

I've reviewed the code changes; I think having the ability to have multiple (semi-)concurrent outputs based on config variations is a worthwhile addition.

I don't think I have strong opinions about the implementation specifics, probably because I don't have sufficient experience with using Snakemake in properly complicated workflows.