tree: b01c795d29c600c658fa9c03ddbb8b2c6be20397 [path history] [tgz]
  1. arrow/
  2. Dockerfile

Apache Arrow for Go


Apache Arrow is a cross-language development platform for in-memory data. It specifies a standardized language-independent columnar memory format for flat and hierarchical data, organized for efficient analytic operations on modern hardware. It also provides computational libraries and zero-copy streaming messaging and inter-process communication.

Reference Counting

The library makes use of reference counting so that it can track when memory buffers are no longer used. This allows Arrow to update resource accounting, pool memory such and track overall memory usage as objects are created and released. Types expose two methods to deal with this pattern. The Retain method will increase the reference count by 1 and Release method will reduce the count by 1. Once the reference count of an object is zero, any associated object will be freed. Retain and Release are safe to call from multiple goroutines.

When to call Retain / Release?

  • If you are passed an object and wish to take ownership of it, you must call Retain. You must later pair this with a call to Release when you no longer need the object. “Taking ownership” typically means you wish to access the object outside the scope of the current function call.

  • You own any object you create via functions whose name begins with New or Copy or when receiving an object over a channel. Therefore you must call Release once you no longer need the object.

  • If you send an object over a channel, you must call Retain before sending it as the receiver is assumed to own the object and will later call Release when it no longer needs the object.


The arrow package makes extensive use of c2goasm to leverage LLVM's advanced optimizer and generate PLAN9 assembly functions from C/C++ code. The arrow package can be compiled without these optimizations using the noasm build tag. Alternatively, by configuring an environment variable, it is possible to dynamically configure which architecture optimizations are used at runtime. See the cpu package README for a description of this environment variable.

Example Usage

The following benchmarks demonstrate summing an array of 8192 values using various optimizations.

Disable no architecture optimizations (thus using AVX2):

$ INTEL_DISABLE_EXT=NONE go test -bench=8192 -run=. ./math
goos: darwin
goarch: amd64
BenchmarkFloat64Funcs_Sum_8192-8   	 2000000	       687 ns/op	95375.41 MB/s
BenchmarkInt64Funcs_Sum_8192-8     	 2000000	       719 ns/op	91061.06 MB/s
BenchmarkUint64Funcs_Sum_8192-8    	 2000000	       691 ns/op	94797.29 MB/s
ok	6.444s

NOTE: NONE is simply ignored, thus enabling optimizations for AVX2 and SSE4

Disable AVX2 architecture optimizations:

$ INTEL_DISABLE_EXT=AVX2 go test -bench=8192 -run=. ./math
goos: darwin
goarch: amd64
BenchmarkFloat64Funcs_Sum_8192-8   	 1000000	      1912 ns/op	34263.63 MB/s
BenchmarkInt64Funcs_Sum_8192-8     	 1000000	      1392 ns/op	47065.57 MB/s
BenchmarkUint64Funcs_Sum_8192-8    	 1000000	      1405 ns/op	46636.41 MB/s
ok	4.786s

Disable ALL architecture optimizations, thus using pure Go implementation:

$ INTEL_DISABLE_EXT=ALL go test -bench=8192 -run=. ./math
goos: darwin
goarch: amd64
BenchmarkFloat64Funcs_Sum_8192-8   	  200000	     10285 ns/op	6371.41 MB/s
BenchmarkInt64Funcs_Sum_8192-8     	  500000	      3892 ns/op	16837.37 MB/s
BenchmarkUint64Funcs_Sum_8192-8    	  500000	      3929 ns/op	16680.00 MB/s
ok	6.179s