tarn 0.13.3

A generic framework for key-value storage

A generic framework for key-value storage

Install

pip install tarn

Recipes

A simple datalake

Let's start small and create a simple disk-based datalake. It will store various files, and the keys will be their sha256 digest:

from tarn import HashKeyStorage

storage = HashKeyStorage('/path/to/some/folder')
# here `key` is the sha256 digest
key = storage.write('/path/to/some/file.png')
# now we can use the key to read the file at a later time
with storage.read(key) as value:
    # this will output something like Path('/path/to/some/folder/a0/ff9ae8987..')
    print(value.resolve())

# you can also store values directly from memory
# - either byte strings
key = storage.write(b'my-bytes')
# - or file-like objects
#  in this example we stream data from an url directly to the datalake
import requests

key = storage.write(requests.get('https://example.com').raw)

Smart cache to disk

A really cool feature of tarn is memoization with automatic invalidation:

from tarn import smart_cache


@smart_cache('/path/to/storage')
def my_expensive_function(x):
    y = x ** 2
    return my_other_function(x, y)


def my_other_function(x, y):
    ...
    z = x * y
    return x + y + z

Now the calls to my_expensive_function will be automatically cached to disk.

But that's not all! Let's assume that my_expensive_function and my_other_function are often prone to change, and we would like to invalidate the cache when they do. Just annotate these function with a decorator:

from tarn import smart_cache, mark_unstable


@smart_cache('/path/to/storage')
@mark_unstable
def my_expensive_function(x):
    ...


@mark_unstable
def my_other_function(x, y):
    ...

Now any change to these functions, will cause the cache to invalidate itself!

Other storage locations

We support multiple storage locations out of the box.

Didn't find the location you were looking for? Create an issue.

S3

from tarn import HashKeyStorage, S3

storage = HashKeyStorage(S3('my-storage-url', 'my-bucket'))

Redis

If your files are small, and you want a fast in-memory storage Redis is a great option

from tarn import HashKeyStorage, RedisLocation

storage = HashKeyStorage(RedisLocation('localhost'))

SFTP

from tarn import HashKeyStorage, SFTP

storage = HashKeyStorage(SFTP('myserver', '/path/to/root/folder'))

SCP

from tarn import HashKeyStorage, SCP

storage = HashKeyStorage(SCP('myserver', '/path/to/root/folder'))

Nginx

Nginx has an autoindex option, that allows to serve files and list directory contents. This is useful when you want to access files over http/https:

from tarn import HashKeyStorage, Nginx

storage = HashKeyStorage(Nginx('https://example.com/storage'))

Advanced

Here we'll show more specific (but useful!) use-cases

Fanout

You might have several HDDs, and you may want to keep your datalake on both without creating a RAID array:

from tarn import HashKeyStorage, Fanout

storage = HashKeyStorage(Fanout(
    '/mount/hdd1/lake',
    '/mount/hdd2/lake',
))

Now both disks are used, and we'll start writing to /mount/hdd2/lake after /mount/hdd1/lake becomes full.

You can even use other types of locations:

from tarn import HashKeyStorage, Fanout, S3

storage = HashKeyStorage(Fanout(S3('server1', 'bucket1'), S3('server2', 'bucket2')))

Or mix and match them as you please:

from tarn import HashKeyStorage, Fanout, S3

# write to s3, then start writing to HDD1 after s3 becomes full
storage = HashKeyStorage(Fanout(S3('server2', 'bucket2'), '/mount/hdd1/lake'))

Lazy migration

Let's say you want to seamlessly replicate an old storage to a new location, but copy only the needed files first:

from tarn import HashKeyStorage, Levels

storage = HashKeyStorage(Levels(
    '/mount/new-hdd/lake',
    '/mount/old-hdd/lake',
))

This will create something like a cache hierarchy with copy-on-read behaviour. Each time we read a key, if we don't find it in /mount/new-hdd/lake, we read it from /mount/old-hdd/lake and save a copy to /mount/new-hdd/lake.

Cache levels

The same cache hierarchy logic can be used if you have a combination of HDDs and SSD which will seriously speed up the reading:

from tarn import HashKeyStorage, Levels, Level

storage = HashKeyStorage(Levels(
    Level('/mount/fast-ssd/lake', write=False),
    Level('/mount/slow-hdd/lake', write=False),
    '/mount/slower-nfs/lake',
))

The setup above is similar to the one we use in our lab:

• we have a slow but huge NFS-mounted storage
• a faster but smaller HDD
• and a super fast but even smaller SSD

Now, we only write to the NFS storage, but the data gets lazily replicated to the local HDD and SSD to speed up the reads.

Caching small files to Redis

We can take this approach even further and use ultra fast in-memory storages, such as Redis:

from tarn import HashKeyStorage, Levels, Small, RedisLocation

storage = HashKeyStorage(Levels(
    # max file size = 100KiB
    Small(RedisLocation('my-host'), max_size=100 * 1024),
    '/mount/hdd/lake',
))

Here we use Small - a wrapper that only allows small (<=100KiB in this case) files to be written to it. In our experiments we observed a 10x speedup for reading small files.

Composability

Because all the locations implement the same interface, you can start creating more complex storage logic specifically tailored to your needs. You can make setups as crazy as you want!

from tarn import HashKeyStorage, Levels, Fanout, RedisLocation, Small, S3, SFTP

storage = HashKeyStorage(Levels(
    Small(RedisLocation('my-host'), max_size=10 * 1024 ** 2),
    '/mount/fast-ssd/lake',

    Fanout(
        '/mount/hdd1/lake',
        '/mount/hdd2/lake',
        '/mount/hdd3/lake',

        # nested locations are not a problem!
        Levels(
            # apparently we want mirrored locations here
            '/mount/hdd3/lake',
            '/mount/old-hdd/lake',
        ),
    ),

    '/mount/slower-nfs/lake',

    S3('my-s3-host', 'my-bucket'),

    # pull missing files over sftp when needed
    SFTP('remove-host', '/path/to/remote/folder'),
))

Acknowledgements

Some parts of our cache invalidation machinery were heavily inspired by the cloudpickle project.