docchi

{
  "data0": "oiufsjdsj...", //1 MB of random string
  "data1": "abuisoehg...", //1 MB of random string
  //...
  "data9": "bhsiofdis...", //1 MB of random string
}

Pseudo-Code

for i in 0..100{
    modify_one_string(json.get_data_mut(rand.gen_range(0..10)));
    let s = serde_json::to_string(&json);
    let mut file = std::fs::File::create(format!("{}/d{}.json", json_dir, i))?;
    file.write_all(s.as_bytes())?;
}

The JSON has ten random 1 MB strings, so the entire JSON file is about 10 MB.

We modified one string and saved as JSON format at a time, and it was repeated 100 times. The total amount of the files was about 1 GB.

It means 10 % of the data is modified each time. Data tends to be modified partially, so 10 % is not very uncommon setting, I think.

Equivalent Docchi data is created, modified, and saved 100 times.

Docchi saves "diff". Only 1 MB of the modified data is saved at best.

The result is below

JSON
sum of file sizes 1021435957
1906 milliseconds

Docchi
sum of file sizes 173005171
604 milliseconds

JSON saved about 1 GB of data and took 1906 milliseconds.

Docchi saved about 173 MB of data and took 604 milliseconds.

Docchi took 17 % of the storage space and about one-third of the time.

Very nice?

Docchi only saved 17 % of the data so of course it's faster.

For comparison, we changed JSON strings to the length of 17 % and run the demo.

JSON(short)
sum of file sizes 173570901
338 milliseconds

About the same file size, and JSON was twice as fast as Docchi.

Serde is very fast, so the result is comprehensible.

But I think Docchi's overhead is reasonable, and Docchi can save in non-blocking manner, so saving time may not worsen user experience.

*Load Demo

Loading is where Docchi pays the cost. Docchi creates "Diff Hierarchy".

Diff Hierarchy Concept

Diff0(10 MB) - Diff00(1 MB) - Diff000(1 MB)
                                   │
                              Diff001(1 MB)
                                   │
                              Diff002(1 MB)
             - Diff01(5 MB) - Diff010(1 MB)
                                   │
                              Diff011(1 MB)
                                   │
                              Diff012(1 MB)
                                  ...
             - Diff02(10 MB)- Diff020(1 MB)
                                   │
                              Diff021(1 MB)
                                  ...
                  ...
 Diff1(10 MB)
   ...

To load Docchi's diffs, we must load files hierarchically from top to bottom, and apply diffs repeatedly.

We used the default setting of Docchi, and it takes 13 files to load one data at most.

The total file size to load can be 4 times bigger than the biggest diff file (10 MB in this case), so it's 40 MB.

We searched the deepest file from the hierarchy and loaded the data.

Docchi 
40 milliseconds

JSON
94 milliseconds

JSON(Short)
16 milliseconds

The Docchi's total amount of data is 4 times bigger than JSON's, but more than twice as fast as JSON.

Docchi is a binary data format and efficiently multi-threaded for loading, so it was able to beat Serde, I think.

*How can it be done?

Constructing diff is very costly process in nature, but Rust's Arc(Atomic-Reference-Count-Pointer) makes it very easy.

Docchi's data is cloned on saving, so non-blocking saving can be done. Docchi's data consists of Arcs, so the cloning can be done instantly.

Using Arc::make_mut, actual copy of the inner data happens when two different Arcs point to the same object, and one of them is modified. When it's modified, two Arcs point to the different objects, so comparing two pointers of Arcs is enough to confirm if it's modified. Comparing actual values is not necessary.

And actual copy happens on the part which is to be actually modified. Copying everything is also unnecessary. Rust's Arc is really magical.

When comparing data and constructing diff, we compare the current object with the cloned object on save, and compare pointers. it's very fast process.

On the other hand, we didn't do anything special on loading. If it's fast, it owes to Rayon.

License

Licensed under either of

Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your discretion.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.

I'm eagerly looking for someone who can

correct my English(I'm not an English speaker)
criticize my code/API

Dependencies

~6–8MB
~150K SLoC