Comparison with Angstrom

How Parseff compares to Angstrom in performance, API style, and trade-offs.

Angstrom is the most widely used parser combinator library in the OCaml ecosystem. This page compares Parseff and Angstrom side by side: performance, API style, and when to use each.

Performance

This page focuses on the cross-library parser workloads in bench/: JSON, CSV, and arithmetic. As a high-level snapshot, the strongest Parseff results in those workloads are:

Benchmark    Best Parseff result                     What it measures
JSON         ~2.30M/s optimized                      Full JSON array parser vs Angstrom
CSV          ~3.62M/s fused                          CSV field parser vs Angstrom and MParser
Arithmetic   ~9.67M/s fused                          Expression parser/evaluator vs Angstrom and MParser

Those rows are useful for a high-level snapshot, but they are not directly comparable across rows because the workloads differ.

JSON benchmark

Benchmarked on a JSON array parser ({[1, 2, 3, ..., 10]}) over 1,000,000 iterations (3 runs). Source: bench/bench_json.ml.

                        Parses/sec     vs. Angstrom (generic)   Minor allocs
Parseff (optimized)     ~2,300,000     1.9x faster              1.3 GB
Parseff (generic)       ~1,560,000     1.3x faster              800 MB
Angstrom (optimized)    ~1,550,000     1.3x faster             11.1 GB
Angstrom (generic)      ~1,190,000     baseline                 5.9 GB

Optimized uses sep_by_take_span plus List.map over zero-copy spans. Scanning avoids per-element String.sub allocation during parsing, while value conversion still uses the stdlib float_of_string via Parseff.span_to_string.

Generic uses the ordinary char, skip_while, take_while, and sep_by combinators with the same float_of_string conversion as Angstrom (generic), so it reflects the baseline direct-style API.

All parsers produce the same output (float list) from the same input and require full input consumption.

Minor allocation totals are the cumulative values reported by Benchmark.latencyN for the full 1,000,000-parse run.

Against each library's best JSON parser, Parseff's optimized path is still ~1.5x faster, while the generic Parseff parser is roughly on par with Angstrom's optimized parser and still ~1.3x faster than Angstrom's generic baseline.

Other cross-library benchmarks

The other cross-library workloads in bench/ show the same pattern:

                       Parseff best     Parseff generic   Angstrom best   MParser
CSV                    ~3,620,000       ~2,500,000       ~2,290,000      ~1,340,000
Arithmetic             ~9,670,000       ~2,080,000       ~4,490,000      ~1,170,000

On CSV, Parseff's generic parser is still ~9% faster than Angstrom's best result, and the fused path is ~1.6x faster. On arithmetic, Parseff's generic parser is ~1.5x faster than Angstrom's generic baseline, and the fused evaluator is ~2.2x faster than Angstrom's optimized parser.

Why Parseff is faster

Direct character scanning. Parseff.take_while runs a tight while loop with character predicates. No regex compilation, no automaton overhead.

Fewer allocations. Span-based APIs return { buf; off; len } slices of the input string without calling String.sub. Angstrom's take_while1 allocates a new string per call.

Fused operations. Parseff.sep_by_take_span parses an entire separated list in one fused operation. Angstrom's equivalent chains sep_by, char, skip_while, and take_while1 through monadic operators, each creating closures.

No monadic overhead. Parsers are direct function calls. No CPS, no closure allocation for sequencing.

API style

The fundamental difference: Parseff uses direct-style imperative code. Angstrom uses monadic composition.

Sequencing

Parseff:

let key_value () =
  let key = Parseff.take_while ~at_least:1 (fun c -> c <> ':') ~label:"key" in
  let _ = Parseff.char ':' in
  Parseff.skip_whitespace ();
  let value = Parseff.take_while ~at_least:1 (fun c -> c <> '\n') ~label:"value" in
  (key, value)

Angstrom:

let key_value =
  take_while1 (fun c -> c <> ':') >>= fun key ->
  char ':' >>= fun _ ->
  skip_while is_ws >>= fun () ->
  take_while1 (fun c -> c <> '\n') >>= fun value ->
  return (key, value)

Both do the same thing. Parseff reads like sequential OCaml code. Angstrom threads results through >>= and return.

Alternation

Parseff:

let value () =
  Parseff.one_of
    [ null_parser; bool_parser; number_parser; string_parser ]
    ()

Angstrom:

let value =
  null_parser <|> bool_parser <|> number_parser <|> string_parser

Similar readability. Angstrom's <|> is more concise. Parseff's Parseff.one_of is explicit about the list structure.

Repetition

Parseff:

let numbers () =
  Parseff.sep_by
    (fun () ->
      Parseff.skip_whitespace ();
      let s = Parseff.take_while ~at_least:1 is_digit ~label:"digit" in
      Parseff.skip_whitespace ();
      int_of_string s)
    (fun () -> Parseff.char ',')
    ()

Angstrom:

let numbers =
  sep_by (ws *> char ',' <* ws)
    (take_while1 is_digit >>| int_of_string)

Angstrom is more concise here thanks to applicative operators (*>, <*). Parseff is more explicit: whitespace handling is visible, not hidden in operator chains.

A complete side-by-side

Here's the same expression parser in both libraries:

Parseff:

let rec expr () =
  Parseff.fold_left
    term
    (fun () ->
      Parseff.skip_whitespace ();
      let _ = Parseff.char '+' in
      Parseff.skip_whitespace ();
      fun a b -> a + b)
    ()

and term () =
  Parseff.fold_left
    factor
    (fun () ->
      Parseff.skip_whitespace ();
      let _ = Parseff.char '*' in
      Parseff.skip_whitespace ();
      fun a b -> a * b)
    ()

and factor () =
  Parseff.or_
    (fun () ->
      let _ = Parseff.char '(' in
      let e = expr () in
      let _ = Parseff.char ')' in
      e)
    (fun () -> Parseff.digit ())
    ()

Angstrom:

let expr =
  fix (fun expr ->
    let factor =
      char '(' *> expr <* char ')'
      <|> (satisfy is_digit >>| fun c -> Char.code c - 48)
    in
    let term =
      chainl1 factor (ws *> char '*' <* ws >>| fun _ -> ( * ))
    in
    chainl1 term (ws *> char '+' <* ws >>| fun _ -> ( + ))
  )

Angstrom is denser. Parseff is more readable for people who aren't fluent in monadic/applicative operators.

Feature comparison

Broader comparison

• Note

  MParser and Opal require explicit backtracking (like Parsec's try). Angstrom and Parseff backtrack automatically on alternation. MParser and Opal don't support streaming input. Only Parseff supports custom typed errors beyond strings.