Benchmarks

Two benchmarks are included: one measuring the cost of transpilation (a one-time import-time cost), and one measuring the cost of calling a transpiled function at runtime.

All results were measured on a single machine and will vary by hardware. Run the benchmarks yourself to get numbers for your environment.

Transpilation

Measures how long transpile() takes to convert a fragments source file into Python. This cost is paid once per file at import time, not on every request.

Source: benchmarks/transpilation/sample.py: a realistic file with two components, a for loop, interpolations, and an HTML comment.

python benchmarks/transpilation/benchmark.py

Transpiler benchmark (5000 iterations × 5 repeats)
  Fragments: 144.7 µs
  Jinja2:    1039.7 µs

Fragments transpiles a file like this in 144.7 µs — about 7× faster than Jinja2 compiles an equivalent template. At 6,900 transpilations per second, a file can be transpiled 1,200 times per character typed at a fast typing speed — this will never be the limiting factor in LSP or IDE performance.

Calls

Measures how long a transpiled fragment function takes to execute at runtime compared to an equivalent plain string interpolation. This is the cost paid on every request.

Source: benchmarks/calls/sample.py: three functions that each render a list of three posts with a heading, byline, and summary: one using fragments, one using Jinja2, and one using str.format().

python benchmarks/calls/benchmark.py

Call benchmark (10000 iterations × 5 repeats)
  Fragments:            2.1 µs
  Jinja2:               5.9 µs
  String interpolation: 0.9 µs

Fragments adds a small overhead over raw string interpolation and is significantly faster than Jinja2. For a typical web request this cost is negligible: at 2.1 µs per render, a single millisecond of request budget can fit ~476 fragment renders.