Performance optimizations - Dedicated functions for map and list diffs

[Valian]

Hi 👋 Thanks for an amazing library. While working on LiveVue patches, I've spotted some low-hanging fruits about performance of this package. So here I go 😄

Approach

First, I've created a small benchmark in test/benchmark.exs using Benchee (I've added it to mix.exs, only in dev). Then I've developed a new version of diff alongside existing one, and after being done I've replaced original code to create this PR.

You can start benchmark by running mix run test/benchmark.exs, but currently it will only show numbers for existing implementation. To compare it to the old one we'd need to copy an old implementation into a separate module.

Overview of changes

My goal was to reduce the number of memory allocations. Each time I could skip creating a new object, I tried to do so.

It turns out that a dedicated function for making a diff for maps and for lists is the best way, as it allows to write a highly-optimized code that almost entirely avoids memory allocations.

Results

In my tests new implementation is 2-3x faster, and consumes around 2.5 less memory. Log from benchee:

Benchmarking Faster JsonPatch with input Complex Lists - Task Management ...
Benchmarking Faster JsonPatch with input Complex Maps - E-commerce Order ...
Benchmarking Faster JsonPatch with input Deep Nesting - Configuration Tree ...
Benchmarking Faster JsonPatch with input Mixed Maps and Lists - Social Media Post ...
Benchmarking Original JsonPatch with input Complex Lists - Task Management ...
Benchmarking Original JsonPatch with input Complex Maps - E-commerce Order ...
Benchmarking Original JsonPatch with input Deep Nesting - Configuration Tree ...
Benchmarking Original JsonPatch with input Mixed Maps and Lists - Social Media Post ...
Calculating statistics...
Formatting results...

##### With input Complex Lists - Task Management #####
Name                         ips        average  deviation         median         99th %
Faster JsonPatch         93.72 K       10.67 μs    ±34.80%        9.67 μs       22.33 μs
Original JsonPatch       34.20 K       29.24 μs   ±315.50%       25.58 μs       69.88 μs

Comparison: 
Faster JsonPatch         93.72 K
Original JsonPatch       34.20 K - 2.74x slower +18.57 μs

Memory usage statistics:

Name                  Memory usage
Faster JsonPatch           8.27 KB
Original JsonPatch        21.97 KB - 2.66x memory usage +13.70 KB

**All measurements for memory usage were the same**

Reduction count statistics:

Name               Reduction count
Faster JsonPatch            1.10 K
Original JsonPatch          3.00 K - 2.73x reduction count +1.91 K

**All measurements for reduction count were the same**

##### With input Complex Maps - E-commerce Order #####
Name                         ips        average  deviation         median         99th %
Faster JsonPatch        100.00 K       10.00 μs   ±139.63%        8.75 μs       26.58 μs
Original JsonPatch       41.54 K       24.07 μs   ±283.08%       20.92 μs       49.00 μs

Comparison: 
Faster JsonPatch        100.00 K
Original JsonPatch       41.54 K - 2.41x slower +14.07 μs

Memory usage statistics:

Name                  Memory usage
Faster JsonPatch           7.34 KB
Original JsonPatch        16.30 KB - 2.22x memory usage +8.95 KB

**All measurements for memory usage were the same**

Reduction count statistics:

Name               Reduction count
Faster JsonPatch            0.98 K
Original JsonPatch          2.34 K - 2.39x reduction count +1.36 K

**All measurements for reduction count were the same**

##### With input Deep Nesting - Configuration Tree #####
Name                         ips        average  deviation         median         99th %
Faster JsonPatch         60.95 K       16.41 μs    ±26.29%       15.33 μs       28.21 μs
Original JsonPatch       24.94 K       40.10 μs    ±23.62%       38.13 μs       67.61 μs

Comparison: 
Faster JsonPatch         60.95 K
Original JsonPatch       24.94 K - 2.44x slower +23.69 μs

Memory usage statistics:

Name                  Memory usage
Faster JsonPatch          13.27 KB
Original JsonPatch        36.13 KB - 2.72x memory usage +22.86 KB

**All measurements for memory usage were the same**

Reduction count statistics:

Name               Reduction count
Faster JsonPatch            1.77 K
Original JsonPatch          4.70 K - 2.65x reduction count +2.93 K

**All measurements for reduction count were the same**

All tests are passing, so I have a pretty high confidence everything is working in the same way as before. Hope these changes will be useful!

[Valian]

OK found another nice improvement of 50%. It turns out checking if string exists before doing replacement is greatly beneficial for performance. Take a look:

##### With input Complex Lists - Task Management #####
Name                       ips        average  deviation         median         99th %
Faster JsonPatch      144.46 K        6.92 μs    ±53.56%        6.33 μs       21.54 μs
JsonPatch              42.24 K       23.67 μs    ±13.84%       23.50 μs       33.18 μs

Comparison: 
Faster JsonPatch      144.46 K
JsonPatch              42.24 K - 3.42x slower +16.75 μs

##### With input Complex Maps - E-commerce Order #####
Name                       ips        average  deviation         median         99th %
Faster JsonPatch      150.49 K        6.65 μs    ±67.21%        5.75 μs       19.29 μs
JsonPatch              46.75 K       21.39 μs    ±37.90%       20.46 μs       36.01 μs

Comparison: 
Faster JsonPatch      150.49 K
JsonPatch              46.75 K - 3.22x slower +14.74 μs

##### With input Deep Nesting - Configuration Tree #####
Name                       ips        average  deviation         median         99th %
Faster JsonPatch       84.30 K       11.86 μs    ±33.07%       10.58 μs       25.75 μs
JsonPatch              25.54 K       39.15 μs    ±13.86%       37.42 μs       57.54 μs

Comparison: 
Faster JsonPatch       84.30 K
JsonPatch              25.54 K - 3.30x slower +27.29 μs

##### With input Mixed Maps and Lists - Social Media Post #####
Name                       ips        average  deviation         median         99th %
Faster JsonPatch      124.02 K        8.06 μs    ±40.10%        7.08 μs       20.33 μs
JsonPatch              38.22 K       26.16 μs    ±12.99%       25.92 μs       37.50 μs

Comparison: 
Faster JsonPatch      124.02 K
JsonPatch              38.22 K - 3.24x slower +18.10 μs

[corka149]

Hi @Valian thanks for all the effort. <3

I will have a look at it asap.

[corka149]

Looks good to me - please fix just the credo issues.

[Valian]

@corka149 Fixed! 👍🏻