Hashing-function agnostic Cuckoo filters for Redis.
Cuckoo filters are a probabilistic data structure that allows you to test for membership of an element in a set without having to hold the whole set in memory.
This is done at the cost of having a probability of getting a false positive response, which, in other words, means that they can only answer "Definitely no" or "Probably yes". The false positive probability is roughly inversely related to how much memory you are willing to allocate to the filter.
The most iconic data structure used for this kind of task are Bloom filters but Cuckoo filters boast both better practical performance and efficiency, and, more importantly, the ability of deleting elements from the filter.
Bloom filters only support insertion of new items. Some extensions of Bloom filters have the ability of deleting items but they achieve so at the expense of precision or memory usage, resulting in a far worse tradeoff compared to what Cuckoo filters offer.
Cuckoo filters offer a very interesting division of labour between server and clients.
Since Cuckoo filters rely on a single hashing of the original item you want to
insert, it is possible to off-load that part of the computation to the client.
In practical terms it means that instead of sending the whole item to Redis, the
clients send hash
and fingeprint
of the original item.
- You need to push trough the cable a constant amount of data per item instead of N bytes (Redis is a remote service afterall, you're going through a UNIX socket at the very least).
- To perform well, Cuckoo filters rely on a good choice of fingerprint for each item and it should not be left to the library.
- The hash function can be decided by you, meaning that this module is hashing-function agnostic.
The last point is the most important one. It allows you to be more flexible in case you need to reason about item hashes across different clients potentially written in different languages.
Additionally, different hashing function families specialize on different use cases that might interest you or not. For example some work best for small data (< 7 bytes), some the opposite. Some focus more on performance at the expense of more collisions, while some others behave better than the rest on peculiar platforms.
This blogpost shows a few benchmarks of different hashing function families.
Considering all of that, the choice of hashing and fingerprinting functions has to be up to you.
For the internal partial hashing that has to happen when reallocating a fingerprint server-side, this implementation uses FNV1a which is robust and fast for 1 byte inputs (the size of a fingerprint).
Thanks to how Cuckoo filters work, that choice is completely transparent to the clients.
-
Download a precompiled binary from the Release section of this repo or compile it yourself (instructions at the end of this README).
-
Put
libredis-cuckoofilter.so
module in a folder readable by your Redis server. -
To try out the module you can send
MODULE LOAD /path/to/libredis-cuckoofilter.so
using redis-cli or a client of your choice. -
Once you save on disk a key containing a Cuckoo filter you will need to add
loadmodule /path/to/libredis-cuckoofilter.so
to yourredis.conf
, otherwise Redis will not load complaining that it doesn't know how to read some data from the.rdb
file.
redis-cli> MODULE LOAD /path/to/libredis-cuckoofilter.so
OK
redis-cli> CF.INIT test 64K
OK
redis-cli> CF.ADD test 5366164415461427448 97
OK
redis-cli> CF.CHECK test 5366164415461427448 97
(integer) 1
redis-cli> CF.REM test 5366164415461427448 97
OK
redis-cli> CF.CHECK test 5366164415461427448 97
(integer) 0
import redis
r = redis.Redis()
# Load the module if you haven't done so already
r.execute_command("module", "load", "/path/to/libredis-cuckoofilter.so")
# Create a filter
r.execute_command("cf.init", "test", "64k")
# Define a fingerprinting function, for hashing we'll use python's builtin `hash()`
def fingerprint(x):
return ord(x[0]) # takes the first byte and returns its numerical value
item = "banana"
# Add an item to the filter
r.execute_command("cf.add", "test", hash(item), fingerprint(item))
# Check for its presence
r.execute_command("cf.check", "test", hash(item), finterprint(item)) # => true
# Check for a non-existing item
r.execute_command("cf.check", "test", hash("apple"), fingerprint("apple")) # => false
In Cuckoo filters the number of bytes that we decide to use as fingerprint will directly impact the maximum false positive error rate of a given filter. This implementation supports 1, 2 and 4-byte wide fingerprints.
Error % -> 3.125e-02 (~0.03, i.e. 3%)
Error % -> 1.22070312e-04 (~0.0001, i.e. 0.01%))
Error % -> 9.31322574e-10 (~0.000000001, i.e. 0.0000001%)
Returns the correct size for a filter that must hold at most universe
items.
Default fpsize
is 1, specify a different value if you need an error rate lower
than 3%.
Cuckoo filters should never be filled over 80% of their maximum theoretical capacity
both for performance reasons and because a filter that approaces 100% fill rate will
start refusing inserts with a ERR too full
error.
This command will automatically pad universe
for you. Use EXACT
if you don't want
that behavior.
Returns the theoretical maximum number of items that can be added to a filter of given
size
and fpsize
. Default fpsize
is 1.
Instantiates a new filter. Use CF.SIZEFOR
to know the correct value for size
.
Supported sizes are a power of 2 in this range: 1K .. 8G
.
Default error rate is 3%, use fpsize
to specify a different target error rate.
Adds a new item to the filter. Both hash
and fp
must be numbers.
In particular, hash
has to be a 64bit representable number, while fp
should be a fpsize
representable number. As an example, a filter with
fpsize
set to 1
will cause the maximum recommended value of fp
to be 255
.
The fp
argument is a u32
so (2^32)-1
is its maximum valid value, but when
fpsize
is lower than 4
, high bits will be truncated (e.g. -1 == 255
when
fpsize == 1
).
You can use both signed and unsigned values as long as you are consistent
in their use. Internally all values will be transalted to unsigned.
If a filter is undersized/overfilled or you are adding multiple copies of
the same item or, worse, you're not properely handling information entropy,
this command will return ERR too full
.
Read the extented example in
kristoff-it/zig-cuckoofilter
to learn more about misusage scenarios.
Deletes an item. Accepts the same arguments as CF.ADD
.
WARNING: this command must be used to only delete items that were
previously inserted. Trying to delete non-existing items will corrupt the
filter and cause it to lockdown. When that happens all command will start
returning ERR broken
, because at that point it will be impossible to
know what the correct state would be. Incurring in ERR broken
is
a usage error and should never happen. Read the extented example in
kristoff-it/zig-cuckoofilter
to learn more about misusage scenarios.
Checks if an item is present in the filter or not. Returns 1
for the
positive case and 0
otherwise. Accepts the same arguments as CF.ADD
.
Returns the number of items present in the filter.
Returns 1
if the filter was broken because of misusage of CF.REM
,
returns 0
otherwise. A broken filter cannot be fixed and will start
returning ERR broken
from most comamnds.
Returns 1
if the filter is too full, returns 0
otherwise.
This command can return 1
even if you never received a
ERR too full
from a call to CF.ADD
.
Read the extented example in
kristoff-it/zig-cuckoofilter
to learn more about misusage scenarios.
If you are adding and also deleting items from the filter
but in a moment of congestion you ended up ovferfilling the filter,
this command can help re-distribute some items to fix the situation.
It's not a command you should ever rely on because it should never
be needed if you properly sized your filter using CF.SIZEFOR
.
Read the extented example in
kristoff-it/zig-cuckoofilter
to learn more about misusage scenarios.
Checkout kristoff-it/zig-cuckoofilter for more information about advanced usage of Cuckoo filters and how to deal (and most importantly, prevent) failure scenarios.
- Advanced client-side syncrhonization Given that now the logic is bundled in zig-cuckoofilter and that it can now be used by any C ABI compatible target (checkout the repo for examples in C, JS, Python and Go), combined with Streams it would be possible to keep a client-side Cuckoo filter synced with one in Redis, allowing clients to keep reads locally and asyncrhonously sync with Redis to obtain new updates to the filter.
Download the latest Zig compiler version from http://ziglang.org.
$ zig build-lib -dynamic -isystem src --release-fast src/redis-cuckoofilter.zig
$ zig build-lib -dynamic -isystem src --release-fast -target x86_64-linux --library c src/redis-cuckoofilter.zig
Use zig targets
for the complete list of available targets.
MIT License
Copyright (c) 2019 Loris Cro
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.