The LatchFile / LatchDir API is DEPRECATED and will be removed in the future. Please use the LPath API for working with Latch files instead. See here for more details.

The Latch SDK provides a convenient means of referencing files or directories within task functions without worrying about how or when the passed file objects are copied to the task’s machine at execution.

Let’s look at an example.

from pathlib import Path
from latch import small_task
from latch.types import LatchFile, LatchDir
import subprocess

def foo(fastq: LatchFile, output_dir: LatchDir) -> (LatchFile, LatchDir):

    # When you pass parameter values of type LatchFile or LatchDir, the file will
    # be automatically downloaded on whatever machine the task is scheduled on.

    # Passing the parameter value to a python Path object and resolving it is a
    # common pattern to retrieve the full path of the file on the local filesystem for
    # downstream use.

    local_fastq = Path(fastq).resolve()
    local_output_dir = Path(output_dir).resolve()

    # It's now easy to reference the contents of the file in a subprocessed
    # program. Notice how we're 'placing' outputs in a directory we will return.["myprogram", "analyze", "local_fastq", "-o", str(local_output_dir)])

    # We can also simply read out the contents of the file as we would normally.
    with open(local_fastq) as fq:
      reads =

    # Lets make a new file on this machine and return it along with the results of
    # the previous subprocess.
    with open("/root/foobar", "w") as fb:

    # Notice when we return, we must specify *two* values - a local path and a
    # remote path. We need to know where the file is coming from and where it's
    # going. We'll discuss the latch URL scheme in a moment, but just understand
    # it will go back in your filesystem on the LatchBio console for now.
    return LatchFile("/root/foobar", "latch:///foobar.txt"), LatchDir(local_output_dir, output_dir.remote_path)

Writing to an existing remote LatchDir will only add or update files that are in the local LatchDir. It will not affect other files in the LatchDir. The two examples below illustrate how this works: the task updates test.txt without touching foo.txt.

# This task adds test.txt to an existing LatchDir.
def update_dir(
    output_dir: LatchDir,
) -> LatchDir:
    os.mkdir('/root/output') # An empty dir
    os.system('touch /root/output/test.txt') # A file in the dir
    return LatchDir('/root/output', output_dir.remote_path)
Case 1: there is not an existing 'test.txt' file in the LatchDir.
. Original directory
├── foo.txt  [Creation Time: 1/1/2022, 01:00:00 AM]
             [Last Modified: 1/1/2022, 01:00:00 AM]

. New directory
├── foo.txt  [Creation Time: 1/1/2022, 01:00:00 AM]
             [Last Modified: 1/1/2022, 01:00:00 AM] # Note that it doesn't touch foo.txt
├── test.txt [Creation Time: 1/1/2022, 23:00:00 AM]
             [Last Modified: 1/1/2022, 23:00:00 AM]
Case 2: there is an existing 'test.txt' file in the LatchDir.
. Original directory
├── foo.txt  [Creation Time: 1/1/2022, 01:00:00 AM]
             [Last Modified: 1/1/2022, 01:00:00 AM]
├── test.txt [Creation Time: 1/1/2022, 01:00:00 AM]
             [Last Modified: 1/1/2022, 01:00:00 AM]

. New directory
├── foo.txt  [Creation Time: 1/1/2022, 01:00:00 AM]
             [Last Modified: 1/1/2022, 01:00:00 AM]
├── test.txt [Creation Time: 1/1/2022, 01:00:00 AM]
             [Last Modified: 1/1/2022, 23:00:00 AM]

## Local Paths and Remote Paths

In the majority of cases, we can just use a value annotated with `LatchFile` or
`LatchDir` and expect it to yield a file handler pointing to a local file. This
gives good synergy with `Path` or `open` as we've seen above.

However, it is important to understand that these values _really_ have both a
local and remote path associated with them.

# latch/types/

    def local_path(self) -> str:
        """File path local to the environment executing the task."""
        return self._path

    def remote_path(self) -> Optional[str]:
        """A url referencing in object in LatchData or s3."""
        return self._remote_directory

local_path will always be the absolute path on the task’s machine where the file has been copied to (the machine that your code is running on). remote_path will be a remote object URL with s3 or latch as its host.

There are cases when we would want to access these local_path and remote_path attributes directly:

  • Specifying the remote destination of a returned directory (eg. in the above return statement).
  • Manually fetching additional files from s3 similar to a passed file’s remote source.
  • Using the Latch SDK to list other files similar to a passed file (eg. latch ls latch:///foo)

Using Globs to Move Groups of Files

Often times logic is needed to move groups of files together based on a shared pattern. For instance, you may wish to return all files that end with a fastq.gz extension after a trimming task has been run.

To do this in the SDK, you can leverage the file_glob function to construct lists of LatchFiles defined by a pattern.

The class of allowed patterns are defined as globs. It is likely you’ve already used globs in the terminal by using wildcard characters in common commands, eg. ls *.txt.

The second argument must be a valid latch URL pointing to a directory. This will be the remote location of returned LatchFile constructed with this utility.

In this example, all files ending with .fastq.gz in the working directory of the task will be returned to the latch:///fastqc_outputs directory:

from latch.types import file_glob

def task():


    return file_glob("*.fastq.gz", "latch:///fastqc_outputs")

latch:/// URLs

Recall that URLs (Uniform Resource Locators) describe the location of an object on the internet.

A simplified representation of a URL string syntax can be denoted as:


Where and s3://my-bucket/dna.fa are both valid descriptions of objects, a webpage or a fasta file.

When referencing files stored within LatchBio’s managed filesystem (called LatchData) we must use the latch scheme to appropriately resolve objects to the appropriate account.

For instance, latch:///foo.txt might meant two entirely different things in the context of two different accounts. The resolution to retrieve the correct object occurs based on the user that executed the workflow,

Some examples of valid latch URLs referencing objects in a user’s filesystem:

  • latch:///guide_design/off_targets.csv
  • latch:///foo.txt

Note the three slashes. This is not accidental, but is in strict accordance with the URL specification as there is no real user-facing “host” for latch objects.

Shared latch URLs

Paths that are shared amongst accounts will bear the latch://shared/<path> syntax.