pull

Download tracked files or directories from remote storage based on the current dvc.yaml and .dvc files, and make them visible in the workspace.

Synopsis

usage: dvc pull [-h] [-q | -v] [-j <number>] [-r <name>] [-a] [-T]
                [-d] [-f] [-R] [--all-commits] [--run-cache]
                [targets [targets ...]]

positional arguments:
  targets       Limit command scope to these tracked files/directories,
                .dvc files, or stage names.

Description

The dvc push and dvc pull commands are the means for uploading and downloading data to and from remote storage (S3, SSH, GCS, etc.). These commands are similar to git push and git pull, respectively. Data sharing across environments and preserving data versions (input datasets, intermediate results, models, metrics, etc.) remotely are the most common use cases for these commands.

dvc pull downloads tracked data from remote storage to the cache, and links (or copies) the files or directories to the workspace (refer to dvc config cache.type).

Note that pulling data does not affect code, dvc.yaml, or .dvc files. Those should be downloaded with git pull.

It has the same effect as running dvc fetch and dvc checkout:

Tracked files                Commands
---------------- ---------------------------------

remote storage
     +
     |         +------------+
     | - - - - | dvc fetch  | ++
     v         +------------+   +   +----------+
project's cache                  ++ | dvc pull |
     +         +------------+   +   +----------+
     | - - - - |dvc checkout| ++
     |         +------------+
     v
 workspace

The dvc remote used is determined in order, based on

1. the remote fields in the dvc.yaml or .dvc files.
2. the value passed to the --remote option via CLI.
3. the value of the core.remote config option (see dvc remote default).

Without arguments, it downloads all files and directories referenced in the current workspace (found in dvc.yaml and .dvc files) that are missing from the workspace. Any targets given to this command limit what to pull. It accepts paths to tracked files or directories (including paths inside tracked directories), .dvc files, and stage names (found in dvc.yaml).

The --all-branches, --all-tags, and --all-commits options enable pulling files/dirs referenced in multiple Git commits.

After the data is in the cache, dvc pull uses OS-specific mechanisms like reflinks or hardlinks to put it in the workspace, trying to avoid copying. See dvc checkout for more details.

Note that the command dvc status -c can list files referenced in current stages (in dvc.yaml) or .dvc files, but missing from the cache. It can be used to see what files dvc pull would download.

Options

• -a, --all-branches - determines the files to download by examining dvc.yaml and .dvc metafiles in all Git branches, as well as in the workspace. It's useful if branches are used to track experiments. Note that this can be combined with -T below, for example using the -aT flags.

• -T, --all-tags - examines metafiles in all Git tags, as well as in the workspace. Useful if tags are used to mark certain versions of an experiment or project. Note that this can be combined with -a above, for example using the -aT flags.

• -A, --all-commits - examines metafiles in all Git commits, as well as in the workspace. This downloads tracked data for the entire commit history of the project.

• -d, --with-deps - only meaningful when specifying targets. This determines files to pull by resolving all dependencies of the targets: DVC searches backward from the targets in the corresponding pipelines. This will not pull files referenced in later stages than the targets.

• -R, --recursive - determines the files to pull by searching each target directory and its subdirectories for dvc.yaml and .dvc files to inspect. If there are no directories among the targets, this option has no effect.

• -f, --force - does not prompt when removing workspace files, which occurs when these files no longer match the current stages or .dvc files. This option surfaces behavior from the dvc fetch and dvc checkout commands because dvc pull in effect performs those 2 functions in a single command.

• -r <name>, --remote <name> - name of the remote storage to pull from (see dvc remote list).

• --run-cache - downloads all available history of stage runs from the dvc remote (to the cache only, like dvc fetch --run-cache). Note that dvc repro <stage_name> is necessary to checkout these files (into the workspace) and update dvc.lock.

• -j <number>, --jobs <number> - parallelism level for DVC to download data from remote storage. The default value is 4 * cpu_count(). Note that the default value can be set using the jobs config option with dvc remote modify. Using more jobs may speed up the operation.

• -h, --help - prints the usage/help message, and exit.

• -q, --quiet - do not write anything to standard output. Exit with 0 if no problems arise, otherwise 1.

• -v, --verbose - displays detailed tracing information.

Examples

Let's employ a simple workspace with some data, code, ML models, pipeline stages, such as the DVC project created for the Get Started. Then we can see what happens with dvc pull.

$ git clone https://github.com/iterative/example-get-started
$ cd example-get-started

.
├── data
│   └── data.xml.dvc
├── dvc.lock
├── dvc.yaml
...
└── src
    └── <code files here>

We can now just run dvc pull to download the most recent data/data.xml, model.pkl, and other DVC-tracked files into the workspace:

$ dvc pull

$ tree
.
├── data
│   ├── data.xml
│   ├── data.xml.dvc
...
└── model.pkl

We can also download only the outputs of a specific stage:

$ dvc pull train

Example: With dependencies

Delete the .dvc/cache directory first (with rm -Rf .dvc/cache) to follow this example if you tried the previous ones.

Our pipeline has been set up with these stages: prepare, featurize, train, evaluate.

Imagine the remote storage has been modified such that the data in some of these stages should be updated in the workspace.

$ dvc status -c
...
	deleted:            data/features/test.pkl
	deleted:            data/features/train.pkl
	deleted:            model.pkl
	...

One could do a simple dvc pull to get all the data, but what if you only want to retrieve part of the data?

$ dvc pull --with-deps featurize

# Use the partial update...
# Then pull the remaining data:

$ dvc pull
Everything is up to date.

With the first dvc pull we specified a stage in the middle of this pipeline (featurize) while using --with-deps. DVC started with that stage and searched backwards through the pipeline for data files to download. Later we ran dvc pull to download all the remaining data files.

Example: Download from specific remote storage

For using the dvc pull command, a remote storage must be defined. (See dvc remote add.) For an existing project, remotes are usually already set up and you can use dvc remote list to check them. To remember how it's done, and set a context for the example, let's define a default SSH remote:

$ dvc remote add -d r1 ssh://user@example.com/path/to/dvc/remote/storage
$ dvc remote list
r1	ssh://user@example.com/path/to/dvc/remote/storage

DVC supports several remote types.

To download DVC-tracked data from a specific DVC remote, use the --remote (-r) option of dvc pull:

$ dvc pull --remote r1