KISSKissJson

Benchmarks

Measured results. Numbers below were measured with JMH on the hardware and JDK described. Do not generalize without re-running on your own hardware.

Methodology

How to Run

# Full comparison benchmark
mvn -Pbenchmark clean test-compile exec:exec \
  -Djmh.args='JsonLibraryComparisonBenchmark -wi 5 -i 5 -f 2'

# Single benchmark with more precision
mvn -Pbenchmark clean test-compile exec:exec \
  -Djmh.args='JsonLibraryComparisonBenchmark.kissJsonDeserializePojoList100 -wi 10 -i 10 -f 3'

# Normal build (no JMH needed)
mvn -B verify

Jackson is a test-scope dependency activated only in the benchmark profile. It is never a production dependency.

What Is Benchmarked

Benchmark Description POJO Fields
Simple POJO 4 fields: String name, int age, boolean active, double score 4
Nested POJO 3 fields: String name, int value, NestedPojo nested 3
Date POJO 2 fields: LocalDate localDate, Instant instant 2
List of 100 POJOs 100 elements of Simple POJO 400 total
Map Map<String, Object> with mixed value types 4 entries
Escape string Heavy string with control chars (0-31) and Unicode 1

Latest Results

Measured May 1, 2026 on Apple M4, Temurin JDK 21.0.11.

Settings: averageTime, microseconds/op, 5 warmup iterations, 5 measurement iterations, 2 forks.

Serialization

Scenario KissJson (us/op) Jackson (us/op) KissJson / Jackson
Simple POJO 0.169 0.195 0.87x
Nested POJO 0.251 0.199 1.26x
Date POJO 0.307 0.264 1.16x
List of 100 POJOs 12.300 12.491 0.98x
Map 0.208 0.210 0.99x
Escape string 0.608 0.460 1.32x

Deserialization

Scenario KissJson (us/op) Jackson (us/op) KissJson / Jackson
Simple POJO 0.293 0.310 0.94x
Nested POJO 0.431 0.368 1.17x
Date POJO 0.811 0.816 0.99x
List of 100 POJOs 27.080 22.961 1.18x
Map 0.281 0.354 0.79x

Bold ratios below 1.0x indicate KissJson is faster than Jackson.

Optimization History

Pass 3 (May 1, 2026) — Simple POJO Fast Path

Changes:

  1. MethodHandle constructors — ClassModel caches MethodHandle for no-arg constructors. ObjectReader uses MethodHandle.invoke() instead of Constructor.newInstance(), eliminating InvocationTargetException wrapping.
  2. Typed primitive field setters — FieldModel provides setInt, setLong, setBoolean, setDouble, etc. using Field.setXxx directly. Avoids boxing for primitive fields in the fast path.
  3. Simple POJO fast path — readSimplePojoList for POJO lists with default config: zero per-object allocations, no Set tracking, no JsonPath push/restore, no config flag checks per key-value pair, pre-sized ArrayList.
  4. Precomputed class feature flags — ClassModel computes hasAliases, hasRequiredFields, hasDateFields, hasPrimitiveFields, hasNestedObjects at construction time. Enables routing to optimized paths.

Results vs previous pass:

Scenario Before After Delta
Deserialize Simple POJO 0.311 0.293 -5.8%
Deserialize Nested POJO 0.466 0.431 -7.5%
Deserialize PojoList100 29.959 27.080 -9.6%
Deserialize Date POJO 0.866 0.811 -6.4%
Serialize Simple POJO 0.180 0.169 -6.1%
Serialize PojoList100 13.696 12.300 -10.2%
Escape string 0.640 0.608 -5.0%

Pass 2 (April 30, 2026) — FieldType Dispatch and Streaming Tokenizer

Changes:

  1. Streaming tokenizer — Replaced tree-based JsonParser/JsonValue with JsonReader token stream. No intermediate tree for typed reads.
  2. FieldType enum dispatch — FieldModel precomputes field type (STRING, INT, BOOLEAN, DOUBLE, etc.). readFieldValue uses a switch on FieldType instead of a chain of if (type == X.class) checks.
  3. Lazy JsonPath — Mutable Object[] stack instead of StringBuilder. Path string built only on exception.
  4. skipKey() for unknown properties — Avoids String allocation when skipping unknown keys (when checkDuplicates=false).
  5. Direct longValue() for integers — Avoids BigDecimal allocation for integral number fields.

Results vs Pass 1:

Scenario Before After Delta
Deserialize PojoList100 46.382 29.959 -35.4%
Deserialize Simple POJO 0.462 0.311 -32.7%
Deserialize Nested POJO 0.542 0.466 -14.0%

Pass 1 (April 30, 2026) — Initial Baseline

Initial JMH measurements establishing the baseline before optimization work.

Per-Element Cost Breakdown

For PojoList100 deserialization (100 SimplePojo objects, 4 fields each):

Cost Component per element Notes
Object construction ~15 ns MethodHandle.invoke() for no-arg constructor
Field setting (4 fields) ~80 ns Field.setInt/setBoolean/setDouble for primitives, Field.set for String
Token reading (4 keys + 4 values) ~120 ns JsonReader.readKey(), nextToken(), longValue() etc.
HashMap lookup (4 keys) ~30 ns lookupMap.get(key) per field
ArrayList.add ~5 ns Pre-sized, no resizing
Other (loop overhead) ~20 ns hasNextEntry(), nextEntryOrEnd()
Total ~270 ns 27 us for 100 elements

The remaining gap vs Jackson (~23 us) comes from Jackson’s use of bytecode generation for field access, which eliminates the Field.setXxx reflection call overhead entirely.

Environment Notes