Performance
Panko is built for high-throughput serialization. This page reports Panko’s own numbers across a range of serializer shapes, how they scale with collection size and output mode, and where the costs come from. The last section compares them with Panko 0.8.5, the C extension.
All figures below come from the benchmarks/ suite.
How these were measured
- Tooling:
benchmark-ipsfor throughput andmemory_profilerfor allocations, on the same block, after one warm-up call. - Runtime: Ruby 4.0.9 with YJIT enabled, ActiveRecord 8.1, Oj 3.17, SQLite in-memory.
- Machine: Apple M4 Max.
- Rounds: each figure is the median of 12 runs, each in a fresh process.
- Data: records are eager-loaded (associations preloaded), so no N+1 queries happen inside the measured block.
- Collection sizes: 50 and 2,300 records.
- Output modes: JSON (
serialize_to_json/ArraySerializer#to_json) and Hash (serialize/#to_a).
Numbers are indicative of relative cost and will differ on your hardware. Run the suite yourself to get numbers for your environment.
Throughput
Iterations per second, where one iteration serializes the entire collection. Higher is better.
| Shape | Records | serialize_to_json | serialize (Hash) |
|---|---|---|---|
| Simple: 5 flat attributes | 50 | 74.4K i/s | 81.0K i/s |
| 2,300 | 1.65K i/s | 1.81K i/s | |
| Method attribute | 50 | 115.1K i/s | 116.5K i/s |
| 2,300 | 2.50K i/s | 2.60K i/s | |
has_one association | 50 | 54.9K i/s | 56.3K i/s |
| 2,300 | 1.18K i/s | 1.21K i/s | |
has_many: 5 children each | 50 | 20.9K i/s | 18.4K i/s |
| 2,300 | 416 i/s | 385 i/s | |
| Wide: 70 attributes | 50 | 3.93K i/s | 3.98K i/s |
| 2,300 | 81.0 i/s | 79.3 i/s |
Throughput scales close to linearly with the number of records, so the per-record cost stays roughly constant across collection sizes. For the simple shape in JSON, that’s about 3.7M records per second at 50 records and 3.8M at 2,300.
Allocations
Objects allocated per call. Fewer is better: allocations drive garbage collection, which is a large part of serialization cost.
| Shape | Records | serialize_to_json | serialize (Hash) |
|---|---|---|---|
| Simple | 50 | 3 | 53 |
| 2,300 | 3 | 2,303 | |
| Method attribute | 50 | 3 | 53 |
| 2,300 | 3 | 2,303 | |
has_one | 50 | 3 | 103 |
| 2,300 | 3 | 4,603 | |
has_many | 50 | 3 | 353 |
| 2,300 | 3 | 16,103 | |
| Wide (70 attrs, 5 decimals and 5 dates) | 50 | 503 | 553 |
| 2,300 | 23,006 | 25,303 |
Two things stand out:
-
JSON output allocates a constant handful of objects, no matter how many records: 3, whether you serialize 50 or 2,300. Panko streams values straight into an
Oj::StringWriterrather than building an intermediate Hash, so the collection size doesn’t register in allocations. This holds for strings, integers, booleans, method attributes, and associations. -
Hash output allocates one Hash per serialized object (each record plus each associated record), so its allocation count grows with the collection. That’s inherent to returning a materialized Hash. Use JSON output when you’re producing a JSON response and don’t need the Hash.
The Wide row is the exception. Each decimal and date is formatted into a String, and formatting allocates: 10 objects per record for its 5 decimals and 5 dates, in both modes. Hash output adds its one Hash per record on top.
Compared with 0.8.5
The same scenarios, run against Panko 0.8.5 (the C extension) on the same Ruby and machine, alternating the two engines in each round. Both produced identical output on every row before timing.
| Output | Speed vs 0.8.5 | Allocations vs 0.8.5 |
|---|---|---|
| JSON, 38 rows | 1.78x faster (0.91x to 3.29x) | 3.9x fewer |
| Hash, 26 rows | 2.88x faster (1.09x to 11.15x) | 1.55x fewer |
Figures are geometric means across rows, with the slowest and fastest row in parentheses. Two cases do not win:
- JSON output of a
jsonorjsonbcolumn runs at parity with 0.8.5 (1.01x to 1.04x) and allocates 3 objects per record where 0.8.5 allocated 1. - JSON output of date and time values is 0.91x of 0.8.5 for 50 records. Hash output of the same data is 1.09x.
Reproducing these numbers
Each scenario is a standalone script under benchmarks/. The scripts need an appraisal gemfile (for ActiveRecord and the comparison libraries), so run them through Appraisal rather than bare ruby:
bundle exec appraisal 8.1.0 ruby benchmarks/simple.rb
bundle exec appraisal 8.1.0 ruby benchmarks/has_one.rb
bundle exec appraisal 8.1.0 ruby benchmarks/has_many.rb
bundle exec appraisal 8.1.0 ruby benchmarks/method_attribute.rb
bundle exec appraisal 8.1.0 ruby benchmarks/wide_attributes.rb
8.1.0 is the appraisal name; 7.2.0 and 8.0.0 are also available. A different Ruby or Rails version shifts the absolute figures.
Useful environment knobs:
SIZE=n: run a single collection size instead of the default 50 and 2,300.TARGET=<substr>: run only rows whose label contains it (e.g.TARGET=panko/).PROFILE=memory: print a fullmemory_profilerbreakdown per row.IPS_TIME=s/IPS_WARMUP=s: measured and warm-up seconds per row (default 5 and 2).