Awesome

BSONEach

This module aims on reading large BSON files with low memory consumption. It provides single BSONEach.each(func) function that will read BSON file and apply callback function func to each parsed document.

File is read by 4096 byte chunks, BSONEach iterates over all documents till the end of file is reached.

Also you can use BSONEach.stream(path) if you want to read file as IO stream, which is useful when you use GenStage behavior.

Performance

• This module archives low memory usage (on my test environment it's constantly consumes 28.1 Mb on a 1.47 GB fixture with 1 000 000 BSON documents).

• Correlation between file size and parse time is linear. (You can check it by running mix bench).

  $ mix bench
  Settings:
    duration:      1.0 s
  
  ## IterativeBench
  [17:36:14] 1/8: read and iterate 1 document
  [17:36:16] 2/8: read and iterate 30 documents
  [17:36:18] 3/8: read and iterate 300 documents
  [17:36:20] 4/8: read and iterate 30_000 documents
  [17:36:21] 5/8: read and iterate 3_000 documents
  ## StreamBench
  [17:36:22] 6/8: stream and iterate 300 documents
  [17:36:24] 7/8: stream and iterate 30_000 documents
  [17:36:25] 8/8: stream and iterate 3_000 documents
  
  Finished in 13.19 seconds
  
  ## IterativeBench
  benchmark name                       iterations   average time
  read and iterate 1 document              100000   15.54 µs/op
  read and iterate 30 documents             50000   22.63 µs/op
  read and iterate 300 documents              100   13672.39 µs/op
  read and iterate 3_000 documents             10   127238.70 µs/op
  read and iterate 30_000 documents             1   1303975.00 µs/op
  ## StreamBench
  benchmark name                       iterations   average time
  stream and iterate 300 documents            100   14111.38 µs/op
  stream and iterate 3_000 documents           10   142093.60 µs/op
  stream and iterate 30_000 documents           1   1429789.00 µs/op
  

• It's better to pass a file to BSONEach instead of stream, since streamed implementation works so much slower.

• BSONEach is CPU-bounded. Consumes 98% of CPU resources on my test environment.

• (time is not a best way to test this, but..) on large files BSONEach works almost 2 times faster comparing to loading whole file in memory and iterating over it:

  Generate a fixture:

  $ mix generate_fixture 1000000 test/fixtures/1000000.bson
  

  Run different task types:

  $ time mix count_read test/fixtures/1000000.bson
  Compiling 2 files (.ex)
  "Done parsing 1000000 documents."
  mix print_read test/fixtures/1000000.bson  59.95s user 5.69s system 99% cpu 1:05.74 total
  

  $ time mix count_stream test/fixtures/1000000.bson
  Compiling 2 files (.ex)
  Generated bsoneach app
  "Done parsing 1000000 documents."
  mix count_stream test/fixtures/1000000.bson  45.37s user 2.74s system 102% cpu 46.876 total
  

• This implementation works faster than timkuijsten/node-bson-stream NPM package (we comparing with Node.js on file with 30k documents):

  $ time mix count_stream test/fixtures/30000.bson
  "Done parsing 30000 documents."
  mix count_stream test/fixtures/30000.bson  1.75s user 0.35s system 114% cpu 1.839 total
  

  $ time node index.js
  Read 30000 documents.
  node index.js  2.09s user 0.05s system 100% cpu 2.139 total
  

Installation

It's available on hex.pm and can be installed as project dependency:

1. Add bsoneach to your list of dependencies in mix.exs:

```elixir

def deps do
  [{:bsoneach, "~> 0.4.1"}]
end
```

2. Ensure bsoneach is started before your application:

```elixir
def application do
  [applications: [:bsoneach]]
end
```

How to use

1. Open file and pass iostream to a BSONEach.each(func) function:

```elixir

"test/fixtures/300.bson" # File path
|> BSONEach.File.open # Open file in :binary, :raw, :read_ahead modes
|> BSONEach.each(&process_bson_document/1) # Send IO.device to BSONEach.each function and pass a callback
|> File.close # Don't forget to close referenced file
```

2. Callback function should receive a struct:

```elixir
def process_bson_document(%{} = document) do
  # Do stuff with a document
  IO.inspect document
end
```

When you process large files its a good thing to process documents asynchronously, you can find more info here.

Thanks

I want to thank to @ericmj for his MongoDB driver. All code that encodes and decodes to with BSON was taken from his repo.