One of the core aims of orderly2 is to allow collaborative analysis; to do this the end of one piece of work is an input for another piece of work, perhaps someone else’s. To make this work in practice, one orderly2 report can “depend” on some completed packet (or several completed packets) in order to pull in files as inputs.

There are two levels that it is useful to think about dependencies:

At the level of the source report, as an instruction about what we intend to depend on; this may or may not be satisfiable
At the level of a completed packet, as a record about what was depended on

This perspective differs somewhat from workflow managers where it is common to talk about “outdated dependencies” and have some single idea of an end result that a chain of dependencies builds up to.

This vignette walks through some of the practical issues around creating and working with dependencies between reports, starting from simple cases (these will be familiar to users of orderly1) through to more advanced cases. We then cover how to interrogate the dependency graph and our ideas for extending this in future, and some practical issues around how dependencies interact with different locations (there is some overlap here with vignette("collaboration"), which we will highlight).

Using dependencies

Here, we show how to practically use dependencies in a few common scenarios of increasing complexity. The code examples are purposefully too-simple in order to keep the presentation straightforward, see the end of this document for a discussion of how complex these pieces of code might “optimally” be.

Basic use

The primary mechanism for using dependencies is to call orderly2::orderly_dependency() from within an orderly file; this finds a suitable completed packet and copies files that are found from within that packet into your current report.

## src
## ├── analysis
## │   └── analysis.R
## └── data
##     ├── data.R
##     └── data.csv

and src/analysis/analysis.R contains:

orderly2::orderly_dependency("data", "latest()", "data.rds")
d <- readRDS("data.rds")
png("analysis.png")
plot(y ~ x, d)
dev.off()

Here, we’ve used orderly2::orderly_dependency() to pull in the file data.rds from the most recent version (latest()) of the data packet, then we’ve used that file as normal to make a plot, which we’ve saved as analysis.png (this is very similar to the example from vignette("introduction"), to get us started).

id1 <- orderly2::orderly_run("data")
## ℹ Starting packet 'data' `20241106-132351-2348286e` at 2024-11-06 13:23:51.142631
## > d <- read.csv("data.csv")
## > d$z <- resid(lm(y ~ x, d))
## > saveRDS(d, "data.rds")
## ✔ Finished running 'data.R'
## ℹ Finished 20241106-132351-2348286e at 2024-11-06 13:23:51.16383 (0.02119875 secs)
id2 <- orderly2::orderly_run("analysis")
## ℹ Starting packet 'analysis' `20241106-132351-305ed356` at 2024-11-06 13:23:51.193696
## > orderly2::orderly_dependency("data", "latest()", "data.rds")
## ℹ Depending on data @ `20241106-132351-2348286e` (via latest(name == "data"))
## > d <- readRDS("data.rds")
## > png("analysis.png")
## > plot(y ~ x, d)
## > dev.off()
## png 
##   2
## ✔ Finished running 'analysis.R'
## ℹ Finished 20241106-132351-305ed356 at 2024-11-06 13:23:51.273079 (0.07938313 secs)

When we look at the metadata for the packet created from the analysis report, we can see it has used 20241106-132351-2348286e as its dependency:

orderly2::orderly_metadata(id2)$depends
##                     packet                  query        files
## 1 20241106-132351-2348286e latest(name == "data") data.rds....

(indeed it had to, there is only one copy of the data packet to pick from).

Filtering candidates by parameters

In the above example, our query was as simple as it could be — the most recently created packet with the name data. One common pattern we see is that an analysis might have a parameter (for example a country name) and a downstream analysis might share that parameter and want to pull in data for a country.

## src
## ├── analysis
## │   └── analysis.R
## └── data
##     └── data.R

with src/data/data.R containing:

orderly2::orderly_parameters(cyl = NULL)
d <- mtcars[mtcars$cyl == cyl, ]
saveRDS(d, "data.rds")

We can run this for several values of cyl:

orderly2::orderly_run("data", list(cyl = 4))
## ℹ Starting packet 'data' `20241106-132351-61f5ea15` at 2024-11-06 13:23:51.387464
## ℹ Parameters:
## • cyl: 4
## > orderly2::orderly_parameters(cyl = NULL)
## > d <- mtcars[mtcars$cyl == cyl, ]
## > saveRDS(d, "data.rds")
## ✔ Finished running 'data.R'
## ℹ Finished 20241106-132351-61f5ea15 at 2024-11-06 13:23:51.416767 (0.02930307 secs)
## [1] "20241106-132351-61f5ea15"
orderly2::orderly_run("data", list(cyl = 6))
## ℹ Starting packet 'data' `20241106-132351-6fc1021e` at 2024-11-06 13:23:51.441139
## ℹ Parameters:
## • cyl: 6
## > orderly2::orderly_parameters(cyl = NULL)
## > d <- mtcars[mtcars$cyl == cyl, ]
## > saveRDS(d, "data.rds")
## ✔ Finished running 'data.R'
## ℹ Finished 20241106-132351-6fc1021e at 2024-11-06 13:23:51.467293 (0.02615428 secs)
## [1] "20241106-132351-6fc1021e"
orderly2::orderly_run("data", list(cyl = 8))
## ℹ Starting packet 'data' `20241106-132351-7da2dcc5` at 2024-11-06 13:23:51.495305
## ℹ Parameters:
## • cyl: 8
## > orderly2::orderly_parameters(cyl = NULL)
## > d <- mtcars[mtcars$cyl == cyl, ]
## > saveRDS(d, "data.rds")
## ✔ Finished running 'data.R'
## ℹ Finished 20241106-132351-7da2dcc5 at 2024-11-06 13:23:51.522356 (0.02705026 secs)
## [1] "20241106-132351-7da2dcc5"

Our follow-on analysis contains:

orderly2::orderly_parameters(cyl = NULL)
orderly2::orderly_dependency(
  "data",
  "latest(parameter:cyl == this:cyl)",
  "data.rds")
d <- readRDS("data.rds")
png("analysis.png")
plot(mpg ~ disp, d)
dev.off()

Here the query latest(parameter:cyl == this:cyl) says “find the most recent packet where it’s parameter”cyl” (parameter:cyl) is the same as the parameter in the currently running report (this:cyl).

orderly2::orderly_run("analysis", list(cyl = 4))
## ℹ Starting packet 'analysis' `20241106-132351-93b7fe71` at 2024-11-06 13:23:51.581846
## ℹ Parameters:
## • cyl: 4
## > orderly2::orderly_parameters(cyl = NULL)
## > orderly2::orderly_dependency(
## +   "data",
## +   "latest(parameter:cyl == this:cyl)",
## +   "data.rds")
## ℹ Depending on data @ `20241106-132351-61f5ea15` (via latest(parameter:cyl == this:cyl && name == "data"))
## > d <- readRDS("data.rds")
## > png("analysis.png")
## > plot(mpg ~ disp, d)
## > dev.off()
## png 
##   2
## ✔ Finished running 'analysis.R'
## ℹ Finished 20241106-132351-93b7fe71 at 2024-11-06 13:23:51.642131 (0.06028557 secs)
## [1] "20241106-132351-93b7fe71"

Interpreting errors

If your query fails to resolve a candidate it will error:

orderly2::orderly_run("analysis", list(cyl = 9000))
## ℹ Starting packet 'analysis' `20241106-132351-adf03924` at 2024-11-06 13:23:51.684201
## ℹ Parameters:
## • cyl: 9000
## > orderly2::orderly_parameters(cyl = NULL)
## > orderly2::orderly_dependency(
## +   "data",
## +   "latest(parameter:cyl == this:cyl)",
## +   "data.rds")
## ✖ Error running 'analysis.R'
## ℹ Finished 20241106-132351-adf03924 at 2024-11-06 13:23:51.744498 (0.06029701 secs)
## Error in `orderly2::orderly_run()`:
## ! Failed to run report
## Caused by error in `outpack_packet_use_dependency()`:
## ! Failed to find packet for query 'latest(parameter:cyl == this:cyl &&
##   name == "data")'
## ℹ See 'rlang::last_error()$explanation' for details

The error message here tries to be fairly self explanatory; we have failed to find a packet that satisfies our querylatest(parameter:cyl == this:cyl && name == "data"); note that the report name data has become part of this query, so there are two conditions being matched on.

The error suggests running rlang::last_error()$explanation for more information, which we can do:

rlang::last_error()$explanation
## Evaluated query: 'latest(A && B)' and found 0 packets
## • A (parameter:cyl == this:cyl): 0 packets
## 
## • B (name == "data"): 3 packets

This is an orderly_query_explain object, which tries to come up with reasons why your query might not have matched; we’ll expand this in the future so let us know what you might like to see.

This tells you that your query can be decomposed into two subqueries A (the match against the parameter cyl being 9000), which matched no packets and B (the match against the packet name being data), which matched 3 packets. If each subquery matched packets but some pairs don’t then it will try and guide you towards problematic pairs.

You can also ask orderly2 to explain any query for you:

orderly2::orderly_query_explain(
  quote(latest(parameter:cyl == 9000)), name = "data")
## Evaluated query: 'latest(A && B)' and found 0 packets
## • A (parameter:cyl == 9000): 0 packets
## 
## • B (name == "data"): 3 packets

If you save this object you can explore it in more detail:

explanation <- orderly2::orderly_query_explain(
  quote(latest(parameter:cyl == 9000)), name = "data")
explanation$parts$B
## $name
## [1] "B"
## 
## $str
## [1] "name == \"data\""
## 
## $expr
## name == "data"
## 
## $n
## [1] 3
## 
## $found
## [1] "20241106-132351-61f5ea15" "20241106-132351-6fc1021e"
## [3] "20241106-132351-7da2dcc5"

(this would have worked with rlang::last_error()$explanation$parts$A too).

You can also use orderly2::orderly_metadata_extract to work out what values you might have looked for:

orderly2::orderly_metadata_extract(
  name = "data",
  extract = c(cyl = "parameters.cyl is number"))
##                         id cyl
## 1 20241106-132351-61f5ea15   4
## 2 20241106-132351-6fc1021e   6
## 3 20241106-132351-7da2dcc5   8

Filtering candidates in other ways

Above we saw two types of filtering candidates: latest() selected the most recent packet while latest(parameter:cyl == this:cyl) found a packet whose parameter matched one of our parameters.

We could have used latest(parameter:cyl == 4) to hard code in a specific parameter value, and used latest(parameter:cyl == environment:cyl) to match against whatever value cyl took in the evaluating environment.

Instead of a query, you can provide a single id (e.g, 20241106-132351-93b7fe71), which would mean that even as new copies of the data packet are created, this dependency will always resolve to the same value.

You can chain together logical operations with && (both sides must be true) or || (either side must be true), and group conditions with parentheses. In addition to ==, the usual complement of comparison operators will work. So you might have complex queries like

latest((parameter:x == 1 || parameter:x == 2) && parameter:y > 10)

but in practice most people have queries that are a series of restrictions with &&.

Computing dependencies and using many dependencies at once

One common pattern is the map-reduce pattern over a set of orderly reports. With this, a set of packets are created over a vector of parameters, or perhaps a chain of different reports for each parameter, then they are all combined together. For some parameter p that takes values “x”, “y” and “z”, this might look like:

    B(p = "x") -- C(p = "x")
  /                          \
A - B(p = "y") -- C(p = "y") - D
  \                          /
    B(p = "z") -- C(p = "z")

So here, D will want to combine all of the three copies of the C packet, one for each of p as “x”, “y” and “z”.

Especially if there are only three values and these are hard coded, you might just write it out as

orderly2::orderly_dependency("C", quote(latest(parameter:p == "x")),
                             c("data/x.rds" = "result.rds"))
orderly2::orderly_dependency("C", quote(latest(parameter:p == "y")),
                             c("data/y.rds" = "result.rds"))
orderly2::orderly_dependency("C", quote(latest(parameter:p == "z")),
                             c("data/z.rds" = "result.rds"))

Note here that in each call we vary the second argument to select a different parameter value, and in the third argument we are naming our destination file a different name (so we end up with three files in data/).

You can write this out as a for loop:

for (p in c("x", "y", "z")) {
  orderly2::orderly_dependency("C", quote(latest(parameter:p == environment:p)),
                               c("data/${p}.rds" = "result.rds"))
}

Here, in the second argument we use environment:p to fetch the value of p from the calling environment - this is the looping value so will take all three values. In the name of the third argument, we use the special interpolation format ${p} to substitute in the value of p to build a filename.

How dependencies interact with locations

By default, any packet that you have unpacked on your local archive is considered a candidate for inclusion by orderly_dependency(). This is not always what you want.

The locations that are selected, and the packets within them that are considered as candidates can be controlled by the search_options argument to orderly2::orderly_run (note that the argument is to orderly_run(), not to orderly_dependency() because this is an effect controlled by the runner of the report, not the writer of the report).

There are three components here that affect how packets are selected

location: is a character vector of locations, matching your location names. Only packets that can be found at these locations will be considered. So if you have a mix of locally created packets as well as ones that other people can see, specifying location = "server" would limit to packets that are available on the server, which means that you will end up with dependencies that you colleagues would also get.
allow_remote: controls if we are willing to download files from a location in order to satisfy a dependency. If TRUE, then when you run the report, it might download files if more recent packets are available on a location than what you have locally.
fetch_metadata: only has an effect if allow_remote is also TRUE; this causes the metadata to be refreshed before dependency resolution.

There is further discussion of the details in ?orderly_run

Other points

If you are used to systems like targets, it is easy to make reports smaller than they need to be. There’s on real need to make these very small, and picking the right size is a challenge.

If they are too small, you’ll end up writing a lot of code to orchestrate running different reports and pulling things together. You’ll end spending a lot of time about whether things are “up to date” with one another because really a group of things always wants to run together.

If they’re too big then you might end up doing more work than you want to do, because in order to make a change to part of a piece of analysis you must run the whole thing again.

Dependencies between packets