Go package containing implementations of encoders and decoders for various data formats.
At Segment, we do a lot of marshaling and unmarshaling of data when sending, queuing, or storing messages. The resources we need to provision on the infrastructure are directly related to the type and amount of data that we are processing. At the scale we operate at, the tools we choose to build programs can have a large impact on the efficiency of our systems. It is important to explore alternative approaches when we reach the limits of the code we use.
This repository includes experiments for Go packages for marshaling and unmarshaling data in various formats. While the focus is on providing a high performance library, we also aim for very low development and maintenance overhead by implementing APIs that can be used as drop-in replacements for the default solutions.
More details about the implementation of this package can be found here.
The json
sub-package provides a re-implementation of the functionalities
offered by the standard library's encoding/json
package, with a focus on lowering the CPU and memory footprint of the code.
The exported API of this package mirrors the standard library's
encoding/json
package, the only
change needed to take advantage of the performance improvements is the import
path of the json
package, from:
import (
"encoding/json"
)
to
import (
"github.com/segmentio/encoding/json"
)
The improvement can be significant for code that heavily relies on serializing and deserializing JSON payloads. The CI pipeline runs benchmarks to compare the performance of the package with the standard library and other popular alternatives; here's an overview of the results (using Go v1.13):
Comparing to encoding/json
goos: linux
goarch: amd64
name old time/op new time/op delta
Marshal/*json.codeResponse2 9.05ms ±12% 6.40ms ±23% -29.34% (p=0.000 n=8+8)
Unmarshal/*json.codeResponse2 35.3ms ± 7% 9.6ms ± 0% -72.83% (p=0.001 n=7+7)
name old speed new speed delta
Marshal/*json.codeResponse2 215MB/s ±13% 310MB/s ±20% +43.80% (p=0.000 n=8+8)
Unmarshal/*json.codeResponse2 55.1MB/s ± 7% 202.5MB/s ± 0% +267.41% (p=0.001 n=7+7)
name old alloc/op new alloc/op delta
Marshal/*json.codeResponse2 0.00B 0.00B ~ (all equal)
Unmarshal/*json.codeResponse2 1.86MB ± 1% 0.01MB ± 1% -99.52% (p=0.000 n=8+8)
name old allocs/op new allocs/op delta
Marshal/*json.codeResponse2 0.00 0.00 ~ (all equal)
Unmarshal/*json.codeResponse2 76.4k ± 0% 0.0k ± 0% -99.95% (p=0.000 n=8+8)
Comparing to github.com/json-iterator/go
goos: linux
goarch: amd64
name old time/op new time/op delta
Marshal/*json.codeResponse2 29.9ms ± 4% 6.4ms ±23% -78.61% (p=0.000 n=7+8)
Unmarshal/*json.codeResponse2 12.6ms ± 6% 9.6ms ± 0% -23.77% (p=0.001 n=7+7)
name old speed new speed delta
Marshal/*json.codeResponse2 64.9MB/s ± 4% 309.8MB/s ±20% +377.19% (p=0.000 n=7+8)
Unmarshal/*json.codeResponse2 152MB/s ±10% 202MB/s ± 0% +32.97% (p=0.000 n=8+7)
name old alloc/op new alloc/op delta
Marshal/*json.codeResponse2 3.40MB ± 0% 0.00MB -100.00% (p=0.000 n=8+8)
Unmarshal/*json.codeResponse2 1.03MB ± 0% 0.01MB ± 1% -99.14% (p=0.001 n=6+8)
name old allocs/op new allocs/op delta
Marshal/*json.codeResponse2 102k ± 0% 0k -100.00% (p=0.000 n=8+8)
Unmarshal/*json.codeResponse2 37.1k ± 0% 0.0k ± 0% -99.89% (p=0.000 n=6+8)
Although this package aims to be a drop-in replacement of encoding/json
,
it does not guarantee the same error messages. It will error in the same cases
as the standard library, but the exact error message may be different.
The iso8601
sub-package exposes APIs to efficiently deal with with string
representations of iso8601 dates.
Data formats like JSON have no syntaxes to represent dates, they are usually
serialized and represented as a string value. In our experience, we often have
to check whether a string value looks like a date, and either construct a
time.Time
by parsing it or simply treat it as a string
. This check can be
done by attempting to parse the value, and if it fails fallback to using the
raw string. Unfortunately, while the happy path for time.Parse
is fairly
efficient, constructing errors is much slower and has a much bigger memory
footprint.
We've developed fast iso8601 validation functions that cause no heap allocations
to remediate this problem. We added a validation step to determine whether
the value is a date representation or a simple string. This reduced CPU and
memory usage by 5% in some programs that were doing time.Parse
calls on very
hot code paths.