MarcusGrass / dir-cache

Low effort sync disk caching lib

dir-cache - A very low-effort directory cache

A map-interface which propagates writes to disk in a somewhat browsable format.

Designed to be simple to use and understand, not particularly effective.

Performance

Bad.

Performance pt. 2

Okay if you write lazily to disk. Since (potentially depending on options) each map operation corresponds to at least multiple
disk-operations, the map is ill-suited to high-frequency operations.

Since a specific value will net a disk-representation that is strictly larger than the raw content size the dir-cache is space-inefficient.

Lastly, file-system specifics may make many keys perform poorly (I'm looking at you NTFS).

Should be used when

You have some toy-project that's pinging an external API and you want to easily just cache responses to certain requests to reduce load on your counterparts servers, or similar.

What you should check out for production use-cases

If you want an embedded KV-store in Rust, consider Sled.

If you want an embedded KV-store not in Rust, consider RocksDB (github), RocksDB(docs.rs).

If you want an embedded SQL-store, not in Rust, consider Sqlite(website), Sync rust crate(Rusqlite), Async rust crate (Sqlx).

Or whatever else.

Why

Now that the above is out of the way, we can get into why this crate exists.

I have in my time exploring public APIs using Rust encountered the same problem many times:

I am exploring an API, taking apart the response and analysing it to inform my further code, but I don't want to make a new web request each iteration out of both latency, and respect for my counterpart.

What I generally have done in these cases is saving the responses to disk and done offline analysis on them. This works, but it's cumbersome, it's handling the same old std::fs::... errors, figuring out a fitting directory structure, and worst of all, writing two separate parts of code, fetch and analyze.

Raison d'etre summarized

I want to write this:

fn iterate_on_api_response_handling(dir_cache: &mut Cache) {
    // This is preferably not dynamic
    let req_key = Path::new("examplerequest");
    // If this has run before then don't send an http request
    let resp = dir_cache.get_or_insert_with(req_key, || {
        let resp = http::client::get("https://example.com")?;
        Ok(resp)
    });
    // Mess around with resp here
    println!("Got length {}", resp.len());
}

With the above, both the fetching and analyzing code can be kept in the same place.

Additionally, if some API returns an unparseable or otherwise unexpected response, it's simple to look at it on disk, fix up the parsing, and then continue on from there, which has been tremendously useful.

Features

The feature set is kept fairly minimal to support the above use case.

Map like interface

There are get, get_or_insert_with, insert, and remove methods on the DirCache .

Browsable disk representation

The values are written to disk at cache-location/{key}/, which makes it easy to check out the saved file, which in my cases are most-often json.

Max age on responses

Since values may become stale, depending on how long the iterating takes, a max age can be set by duration, after which the value will be treated as non-existent. Meaning, running the same get_or_insert_with will the first time fetch data, each time up until the max age has passed, return the cached data, and after the max age has passed fetch new data.

Data optionally saved as generations

Overwriting the same key can optionally shuffle the older key down one generation, leaving it on disk.
Useful in some cases where response changes over time, and you wish to keep a history. Although it's definitely the least useful feature.

Optionally compress generational data

I found some use for this when working with an incredibly sparse json dataset where responses were pretty huge, with the feature lz4 lz4-compression can be picked for old generations.

Caveats

There is one caveat apart from performance that bears consideration.

DISK DANGER

Keys are PathBufs and joined with the dir-cache base directory. This opens up a can of worms, the worst of which is accidentally joining with an abs-path, see the docs on Path::join.

This could potentially lead to destructive results.

There are a few mitigations:

1. Paths are never joined if the right side is absolute, and paths are not allowed to be anything but a Component::Normal. as well as making sure parsed components combined length makes sense with the provided OsStr length (Mitigating unexpected effective paths).
2. Write operations are only done on specific file-names dir-cache-generation-{manifest.txt | n}. (Reducing risk of accidental overwrites of important files).
3. Removal operations are only done on the above specific file-names, as well as empty directories.

This covers all the cases that I can think of, but of course, doesn't cover the cases that I fail to think of.

If using this library, a recommendation is to not use dynamic keys. Fuzzing is done on Linux only, so extra danger if using dynamic keys on other Oses, although it's not safe on Linux just because it's fuzzed.

Async

Mixing this library with async code becomes problematic for two reasons, these will in the end boil down to application performance hits, but again, performance is not considered a priority, so it shouldn't matter, but they're presented below.

Sync disk io

All disk operations are done sync, since regular tokio::fs on Linux discounting tokio-uring, still dispatches synchronous read syscalls, this isn't that much of an issue overall. Where disk the disk-io sync API becomes an issue is if it hogs a lot of time. Alyce Ryhl wrote a great post about why that's problematic a while back.

Get or insert takes an FnOnce, not a Future

Consider this very applicable case for the library, using reqwest:

let key_base = format!("root-to-offset-{offset}");
let key = Path::new(&key_base);
let data = cache.get_or_insert(key, async move {
    let url = format!("{ROOT_URL}&page[offset]={offset}");
    let req = self.inner.get(url).build().unwrap();
    let resp = self.inner.execute(req).await;
    let body = resp.bytes().await.unwrap();
    Ok::<_, Infallible>(body.to_vec())
}).await.unwrap();

That's more or less an excerpt of some code I would have liked to write using this library. reqwest has an async API, I'd like to use it like that.

Instead, I have to do this:

let key_base = format!("root-to-offset-{offset}");
let key = Path::new(&key_base);
let data = cache.get_or_insert(key, || {
    let url = format!("{ROOT_URL}&page[offset]={offset}");
    let req = self.inner.get(url).build().unwrap();
    let resp = self.inner.execute(req);
    let resp = futures::executor::block_on(resp).unwrap();
    let body = futures::executor::block_on(resp.bytes()).unwrap();
    Ok::<_, Infallible>(body.to_vec())
}).unwrap();

I'm forced to use futures to block on the future on the thread currently doing the insert. The benefits of having an async api on reqwest are completely nullified, and my thread is now holding up the executor, causing the same problem as above, possibly to a worse extent.

It's possible to add an async version of get_or_insert easily, but this problem is more theoretical, since performance isn't considered I haven't bothered doing that. Furthermore, I think it would look strange if this specific part of the code consideres async, while the rest uses sync disk-io.

One can fall back to using the plain get, and if that returns None run the async code that generates the data, and insert after, get_or_insert is just a convenience method anyway.

License

The project is licensed under MPL-2.0.