Skip to content
/ ducer Public

Fast and compact maps and sets with Billions of keys for Python, based on finite-state-transducers from the excellent Rust crate fst by Andrew Gallant

License

Notifications You must be signed in to change notification settings

jfolz/ducer

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Ducer documentation

This package provides fast and compact read-only maps and sets that scale up to Billions of keys, while being light on resources with a streamable file format. Complex search patterns that would be infeasible with Python's builtin dict and set are not just possible, but very efficient. All of these amazing things are achieved with finite-state-transducers provided by the excellent Rust crate fst by Andrew Gallant.

Performance

Ducer maps and sets can be built and queried at millions of keys per second. Consider the following example:

import ducer
n = 1_000_000_000
items = ((b"%09d" % i, n-i) for i in range(n))
data = ducer.Map.build(":memory:", items)
m = Map(data)
assert m[b"900000000"] == 100_000_000

In our example, most of the time is spent in Python creating item tuples. Regardless, building happens at almost 4 Million items per second on my humble laptop. Retrieving individual keys is similarly speedy, and simply iterating over all items is twice as fast at 8 Million items per second.

This toy example is almost a best case for FSTs, so the resulting output is just 464 bytes. A real-world example with 1.1 Billion keys, where the key-value pairs occupy 21 GiB without any kind of searchability (stored in already quite compact msgpack format), results in a 4.6 GiB file. Building and retrieval are naturally a bit slower, but still within the 2-3 Million items per second range.

Limitations

Performance is rarely free, so there are some limitations you should consider before proceeding:

  • Keys must be bytes
  • Keys must be inserted in lexicographical order
  • Map values must be non-negative integers less than 2^64
  • Once built, maps and sets cannot be altered

Installation

Most users should be able to simply do:

pip install ducer

To build from source you will need a recent Rust toolchain. Use your preferred method of installation, or follow the official instructions to install Rust. Then run the following at the toplevel of the repository:

pip install .

Building

Above, we already showed that Map.build can build maps in memory by passing ":memory:" as path:

data = ducer.Map.build(":memory:", items)

If you pass any other path (either str or Path; the parent directory must exist), your map will be written directly to that file:

ducer.Map.build("path/to/my.map", items)

Building a map like this uses virtually no extra memory.

Opening

One key advantage of ducer maps is streamability. Unlike the builtin dict, a Map does not have to reside entirely in memory. You can, e.g., use the builtin mmap to stream map data:

with open("path/to/my.map", "rb") as f:
    mm = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ)
m = ducer.Map(mm)

Note that you can safely close the file once the mmap is created. See mmap(2) for details.

Thanks to high compression ratios, pre-loading maps entirely into memory can be feasible. In our experience, at least for local SSD storage, performance is virtually identical, expect pre-loading data takes extra time. Pre-loading can still make sense though, e.g., with networked storage.

with open("path/to/my.map", "rb") as f:
    data = f.read()
m = ducer.Map(data)

Access

To the extent that it's feasible, Map and Set are intended to be direct replacements for the builtin Python dict and set. For m, o: Map and k: bytes, the following works as intended:

k in m
m == o
m[k]
m.get(k)
m.get(k, 42)
len(m)
for k in m:
    pass
for k in m.keys():
    pass
for v in m.values():
    pass
for k, v in m.items():
    pass

For s, o: Set, and k: bytes, the following works as intended:

k in s
s == o
len(s)
for k in s:
    pass
s.isdisjoint(o)
s.issubset(o)
s <= o  # subset
s < o  # proper subset
s.issuperset(o)
s >= o  # superset
s > o  # proper superset

Note: Comparison operations are currently only implemented for other Set objects, not the builtin set. This may change in a future version if there is demand for it.

Differences to builtins

Not implemented

Since Map is immutable, the following are not implemented:

  • clear
  • fromkeys
  • pop
  • popitem
  • setdefault
  • update, |=

Since Set is immutable, the following are not implemented:

  • add
  • clear
  • difference_update, -=
  • discard
  • intersection_update, &=
  • pop
  • remove
  • symmetric_difference_update, ^=
  • update, |=

Further, the |, &, -, ^ operators are also not implemented, since it is not possible to specify the storage path. Use the respective union, intersection, difference, and symmetric_difference methods instead.

Incompatible syntax

difference, intersection, symmetric_difference, and union have slightly different syntax to accomodate the necessary path. For s: Set and others: Iterable[Set]:

s.difference("path/to/result.set", *others)
s.intersection("path/to/result.set", *others)
s.symmetric_difference("path/to/result.set", *others)
s.union("path/to/result.set", *others)

Like the standard library, difference will create the set of all elements of s that are not present in others.

Set operations on maps

Unlike the builtin dict, the ducer Map offers set operations. The syntax is the same as for sets:

m.difference("path/to/result.map", *others)
m.intersection("path/to/result.map", *others)
m.symmetric_difference("path/to/result.map", *others)
m.union("path/to/result.map", *others)

To resolve conflicts between keys present in multiple maps, a list of possible values is assembled. If the key is present in self, then it will be the first value. Values from others are added in given order. By default the last values in the list is used to mimic the behavior of dict.update. Currently, you can choose between these pre-defined operations:

Some examples:

m1 = ducer.Map(ducer.Map.build(":memory:", [(b"k1", 1), (b"k2", 1)]))
m2 = ducer.Map(ducer.Map.build(":memory:", [(b"k2", 2), (b"k3", 2)]))
m3 = ducer.Map(ducer.Map.build(":memory:", [(b"k3", 3)]))
mu = ducer.Map(m1.union(":memory:", m2, m3))
print(dict(mu.items()))
# {b'k1': 1, b'k2': 2, b'k3': 3}
mu = ducer.Map(m1.union(":memory:", m2, m3, select=ducer.Op.First))
print(dict(mu.items()))
# {b'k1': 1, b'k2': 1, b'k3': 2}

Advanced search patterns

The underlying FSTs allow for some advanced search patterns that would otherwise be costly to implement on top of dict and set. Most basic, you can iterate over a range of keys, where ge = greater or equals, gt = greater than, le = less than or equals, and lt = less than:

m.range(ge=b"key17", lt=b"key42")
m.range(gt=b"key17", le=b"key42")

For maps this yields key-value tuples, meaning m.range() without limits is equivalent to m.items().

You can also iterate over all keys that start with a certain prefix, with optional limits same as range:

m.starts_with(b"key", ge=b"key17", lt=b"key42")
m.starts_with(b"key", gt=b"key17", le=b"key42")

You can also search for subsequences , e.g., all keys matching *k*7*, again with optional limits:

m.subsequence(b"k7", ge=b"key17", lt=b"key42")
m.subsequence(b"k7", gt=b"key17", le=b"key42")

Finally, you can create an Automaton to create your own search patterns. The following automata are available:

For example, to recreate Map.starts_with, you can use the following automata with the Map.search method:

a = ducer.Automaton.str(b"key").starts_with()
m.search(a)

Add complement to search for keys that do not start with the string:

a = ducer.Automaton.str(b"key").starts_with().complement()
m.search(a)

Finally, you can combine multiple automata, e.g. with union:

a1 = ducer.Automaton.str(b"key").starts_with()
a2 = ducer.Automaton.str(b"other").starts_with()
a = a1.union(a2)
m.search(a)

Acknowledgements

Ducer is supported by the Albatross and SustainML projects.

About

Fast and compact maps and sets with Billions of keys for Python, based on finite-state-transducers from the excellent Rust crate fst by Andrew Gallant

Topics

Resources

License

Stars

Watchers

Forks