Fast, type-aware data structure cache.
rmemstore
is similar to other caches you may have used, like redis, but it has some differences.
The primary aims of rmemstore
is to be typesafe, fast, and useful as a data structure cache.
Of course, usefulness is an ongoing exercise, as it takes time to grow features. However, rmemstore
is a type-
aware data structure store, which means you can store maps of maps - and the server knows what that means.
It is fast now, however. rmemstore
uses the new Sieve eviction strategy when pressed to eviction. With a 10:1
read:write ratio, 2 threads on an 11 year old Intel i5 server are capable of over 3.3 million operations per
second. Even while being pushed to eviction.
rmemstore
is built on "safe" Rust code. It doesn't rely on subtle tricks to get speed. It does use standard
libraries like the excellent tokio
which may use dark magic, but they're trustworthy.
rmemstore
uses bare tcp - no application frameworks. Each 0 and every 1 that your network card transmits to or
from an rmemstored
server has a direct purpose. Inventing a new ostensibly-portable wire protocol is a vaguely
hubric exercise when suitable alternatives exist. With that in mind, rmemstore
uses protosockets
, which is a
compromise between the aforementioned hubris and pragmatism.
The tcp stream inbound to rmemstored
is a stream of standard, length-delimited protocol buffers rmemstore.Rpc
structures. These messages carry an id, and rmemstored
responds with that id - possibly out of order. It is a
multithreaded, multiplexing server. You can send as much as you want as fast as you can, subject to your network and
cpu capabilities.
The tcp stream outbound from rmemstored
is a stream of standard, length-delimited protocol buffers rmemstore.Response
structures. These messages carry the id from the Rpc that initiated the response. Every rmemstore.Rpc
has a
corresponding rmemstore.Response
.
Inbound and outbound streams are: varint
message
varint
message
[...]. The varint before the message is the
length of the message. So once you have read the bytes for varint
and the length of varint
, you have a complete
message.
You can look at rmem
for an example of how you can use the client. Usage boils down to 3
lines:
let mut configuration = rmemstore::ClientConfiguration::new();
let client = configuration.connect(args.host.to_string()).await?;
client.put("some key", "some value").await?;
You can also put dictionaries:
client.put(
"some key",
HashMap::<&str, &str>::from_iter([
("hello", "world")
]),
).await?;
or dictionaries of strings and dictionaries, however wild you want to get:
client
.put(
"some key",
HashMap::<&str, MemstoreValue>::from_iter([
(
"hello",
MemstoreValue::String {
string: "world".to_string(),
},
),
(
"nested",
MemstoreValue::Map {
map: HashMap::from_iter([(
"inner".to_string(),
MemstoreValue::String {
string: "values".to_string(),
},
)]),
},
),
]),
)
.await?;
You can use rms
to put and get.
For strings, the output is a little more brief.
$ rms put foo `{"string": "some value"}`
$ rms get foo
some value
For maps, the interaction has some verbosity, but it is typed!
$ rms put foo '{"map": {"bar":{"map":{"baz":{"string": "haha"}, "other": {"string": "verbose"}}, "outer": {"string": "another"}}}}'
$ rms get foo
{
"bar": {
"map": {
"baz": {
"string": "haha"
},
"other": {
"string": "verbose"
}
}
}
}
Don't want to use rust? Any tool or language capable of sending and receiving protocol buffers-encoded bytes over
tcp is capable of using rmemstored
. See example-python
for an example in another
language. Note that python, in particular, is a bit of a pain due to not exposing the protobuf varint encoder.
Rather than using the popular moka
cache, rmemstore has its own cache implementation. Here's an example
result from the benchmarks that motivates this deviation:
You can see that the benchmark under eviction favors k-cache at all thread counts. Note that sieve pays on
insert, so this 100% insert benchmark is pessimistic, and get will outperform by a wider margin.