package SZXX

  1. Overview
  2. Docs
Streaming ZIP XML XLSX parser

Install

Dune Dependency

Authors

Maintainers

Sources

4.0.1.tar.gz
md5=7899596c5da8013a459bf72a552eff98
sha512=f2addbe9243c04a7dd21d41972960ecd08c89c9071e1ba0bedac059c198426249c7f3543bead551eb7075403086e9a01c97dcf2263959ecdd978c5e76a4d8235

Description

SZXX is a streaming and efficient XLSX, ZIP and XML parser built from the ground up for low and constant memory usage (<10Mb). SZXX is able to output XLSX rows while reading from a network socket without buffering any part of the file. It can also stream data (including network sockets) out of ZIP and XML files without buffering.

README

SZXX

Streaming ZIP XML XLSX

SZXX is an efficient streaming parser built from the ground up for very low and constant memory usage. It can be used to stream data out of ZIP, XML, and XLSX. It support both files and streams (like HTTP).

There are 3 independent modules. Skip to the one appropriate for your use case.

opam install SZXX

FAQ

SZXX v4 replaces Lwt with Eio. See the release page to upgrade from v3.

SZXX.Xlsx

This example reads from a file and prints a JSON representation of each row:

open! Core
open Eio.Std
open SZXX

let print_xlsx xlsx_path =
  (* The Switch receives any parsing errors *)
  Switch.run @@ fun sw ->
  let file = Eio.Path.(open_in ~sw (Eio.Stdenv.fs env / xlsx_path)) in

  let seq = Xlsx.stream_rows_double_pass ~sw file Xlsx.yojson_cell_parser in

  (* Print each row *)
  Sequence.iter seq ~f:(fun row ->
      `List row.data |> Yojson.Basic.to_string |> print_endline )

Intro to XLSX

XLSX is tabular (like CSV) and typed (like JSON).

Most XLSX files consist of about a dozen separate XML files and folders zipped together.

That's a fair question. SZXX tries extremely hard to quietly handle all the complexity of this chimera of a format.

But the design of the XLSX format demands that you -the user- make a few decisions.

Just the essentials

The types of cells are: String, Formula, Error, Boolean, Number, Date (rarely used), and Null.

Many SZXX.Xlsx functions require a cell_parser argument to convert from XLSX types to your own 'a type.

type 'a cell_parser = {
  string: location -> string -> 'a;
  formula: location -> formula:string -> string -> 'a;
  error: location -> string -> 'a;
  boolean: location -> string -> 'a;
  number: location -> string -> 'a;
  date: location -> string -> 'a;
  null: 'a;
}

This library includes two simple cell_parsers to get started: Xlsx.yojson_cell_parser and Xlsx.string_cell_parser, but creating your own 'a cell_parser is probably a good idea.

XLSX cells use XML-escaping (&amp; for "&", &#x1F600; for "😀" etc). The two aforementioned "beginner" cell_parsers already unescape everything for you. If you write your own cell_parser and if your spreadsheets might contain special characters or non-ASCII characters, you will need to call SZXX.Xml.unescape on data coming from String, Formula, and Error cells.

Most XLSX applications use Number cells (OCaml float) to encode Date and DateTime. Pass this float to Xlsx.parse_date or Xlsx.parse_datetime to decode it. The Date cell type was only introduced to Excel in 2010 and few XLSX files use it.

The vast majority of applications that generate XLSX files will not inline the contents of String cells directly into the spreadsheet. Instead, String cells will contain a reference to an offset in the Shared Strings Table (the SST). This saves space by reusing strings, but 99% of the time those applications will place the SST after the sheet and that's not ideal...

If the XLSX document is in the form of a file, then it's fine because SZXX can jump around the file to parse the SST before reading the rows. But the location of the SST matters when we cannot rewind, such as when parsing directly from an HTTP stream, because it means we might need to cache rows until we can decode them using the SST. As you'll see, SZXX has options to handle this issue.

We can inspect the structure of a file using zipinfo:

zipinfo financial.xlsx

The SST is indicated by xl/sharedStrings.xml and sheets look like xl/worksheets/sheet1.xml.

Xlsx.stream_rows_double_pass

This function extracts rows from an XLSX file. It is guaranteed to run in constant memory, without ever buffering.

It first reads the file's index to locate and extract the SST, then it streams out fully parsed rows from the various sheets. Rows are always emitted strictly in the order they appear in the sheet(s).

open! Core
open Eio.Std
open SZXX

let print_xlsx xlsx_path =
  (* The Switch receives any parsing errors *)
  Switch.run @@ fun sw ->
  let file = Eio.Path.(open_in ~sw (Eio.Stdenv.fs env / xlsx_path)) in

  let seq = Xlsx.stream_rows_double_pass ~sw file Xlsx.yojson_cell_parser in

  (* Print each row *)
  Sequence.iter seq ~f:(fun row ->
      `List row.data |> Yojson.Basic.to_string |> print_endline )

There's just one limitation to stream_rows_double_pass: your XLSX document has to be a file so that SZXX can "rewind" after extracting the SST. To stream XLSX data out of non-seekable streams such as HTTP transfers, use Xlsx.stream_rows_single_pass instead.

Xlsx.stream_rows_double_pass ?filter_sheets ~sw file cell_parser
Arguments:
  • filter_sheets

    • Default: all sheets.

    • Pass a function to limit parsing/extraction to only certain sheets.

    • Sheet IDs start at 1. Note: it does not necessarily match the visual order of the sheets in Excel.

  • sw

    • A regular Eio.Switch.t. This is where parsing errors go.

  • file

    • A file opened with Eio.Path.open_in or Eio.Path.with_open_in.

  • cell_parser

    • A cell parser converts from XLSX types to your own data type (usually a variant).

    • Use Xlsx.string_cell_parser or Xlsx.yojson_cell_parser to get started quickly, then make your own.

Returns:

'a Xlsx.row Sequence.t where 'a is the type of your cell_parser.

Note that a Sequence.t is an ephemeral data structure. Each element can only be seen once, so for example calling Sequence.is_empty will attempt to generate the next element in the Sequence to return true if there was one, but that element will be lost.

SZXX generates elements on demand (lazily), meaning that it will not begin parsing until you begin reading from the Sequence.

Note that you must consume the Sequence from within the Switch. Attempting to read from the Sequence from outside the Switch will fail with Invalid_argument "Coroutine has already failed". The Sequence's contents can safely leave the Switch.

Xlsx.stream_rows_single_pass

This function extracts rows from an XLSX document. Rows encountered before the SST will be buffered (to be streamed out once we've encountered the SST), or dropped (to minimize buffering).

You get full control over any and all buffering.

Xlsx.stream_rows_single_pass ?max_buffering ?filter ?filter_sheets ~sw ~feed cell_parser
Arguments:
  • max_buffering

    • Default: unlimited.

    • Sets a limit to the number of rows that may be buffered while looking for the SST. Raises an exception if it runs out of buffer space before reaching the SST.

  • filter

    • Use this filter to drop uninteresting rows and reduce the number of rows that must be buffered.

    • When filter returns true, the row is buffered

    • When filter returns false, the row is dropped

    • If necessary, use Xlsx.Expert.parse_row_without_sst to access cell-level data. See the examples below.

    • This function is called on every row of every sheet (except for sheets removed by filter_sheets).

  • filter_sheets

    • Default: all sheets.

    • Pass a function to limit parsing/extraction to only certain sheets.

    • Sheet IDs start at 1. Note: it does not necessarily match the visual order of the sheets in Excel.

  • sw

    • A regular Eio.Switch.t. This is where parsing errors go.

  • feed

    • A producer of raw input data. Create a Feed.t by using the SZXX.Feed module.

  • cell_parser

    • A cell parser converts from XLSX types to your own data type (usually a variant).

    • Use Xlsx.string_cell_parser or Xlsx.yojson_cell_parser to get started quickly, then make your own.

Returns:

'a Xlsx.row Sequence.t where 'a is the type of your cell_parser.

Note that a Sequence.t is an ephemeral data structure. Each element can only be seen once, so for example calling Sequence.is_empty will attempt to generate the next element in the Sequence to return true if there was one, but that element will be lost.

SZXX generates elements on demand (lazily), meaning that it will not begin parsing until you begin reading from the Sequence.

Note that you must consume the Sequence from within the Switch. Attempting to read from the Sequence from outside the Switch will fail with Invalid_argument "Coroutine has already failed". The Sequence's contents can safely leave the Switch.

Example 1

Let's read a file in a single pass.

For this example we'll pretend we have these 3 requirements:

  • R1: we're only interested in the first 1000 rows

  • R2: and only the rows where the K column (10th column counting from 0) contains a number >= 100

  • R3: and we only care about sheet #1

We'll use Xlsx.yojson_cell_parser to keep things simple.

First, let's define our filter function. We'll drop all rows that are not useful given the requirements we made up for this example.

let open SZXX in

let filter raw =
  let open Xlsx in
  (* R1: If the row number is over 1000 we don't need to parse the row at all *)
  if raw.row_number > 1000
  then false
  else (
    (* Let's parse the row the best that we can do without having the SST,
       unless we feel like working with raw XML (and we sure don't!) *)
    let row = Expert.parse_row_without_sst yojson_cell_parser raw in

    (* R2: The column at position 10 is the "K" column we're interested in *)
    match List.nth row.data 10 with
    | Some (Available (`Float x)) ->
      (* R2: Keep the row only if the number in "K" is >= 100 *)
      Float.(x >= 100.0)

    | Some (Available _) (* R2: The cell in column "K" isn't a Number *)
    | Some (Delayed _) (* R2: The cell in column "K" is a non-inlined String *)
    | None (* R2: There's too few columns to even have a "K" column *)
      -> false )
in

Our filter function guarantees we'll never buffer more than 1000 rows, so we don't really need to pass ~max_buffering.

(* The Switch receives any parsing errors *)
Switch.run @@ fun sw ->
let src : Eio.Flow.source =
  (* Some kind of data flow such as an HTTP transfer *)
in
let seq =
  Xlsx.stream_rows_single_pass
    ~max_buffering:1000 (* Not necessary but we'll pass this argument anyway just in case we made a mistake in ~filter *)
    ~filter
    ~filter_sheets:(fun ~sheet_id ~raw_size:_ ->
      (* R3: we only care about sheet #1 *)
      sheet_id = 1
    )
    ~sw
    ~feed:(SZXX.Feed.of_flow src) (* See SZXX.Feed for other ways to construct a ~feed *)
    Xlsx.yojson_cell_parser
in

We can now work with the data in seq as usual.

Example 2

For this example let's just count the number of rows across all sheets in the XLSX, with zero buffering.

let open SZXX in
let src : Eio.Flow.source =
  (* Some kind of data flow such as an HTTP transfer *)
in
let count = ref 0 in
let filter _ =
  incr count;
  false (* Always reject *)
in
let seq =
  Xlsx.stream_rows_single_pass
    ~filter
    ~sw
    ~feed:(SZXX.Feed.of_flow src) (* See SZXX.Feed for other ways to construct a ~feed *)
    Xlsx.yojson_cell_parser
in

(* Optional sanity check *)
Sequence.iter seq ~f:(fun _ -> assert false);

(* Print our count! *)
Eio.traceln "Found %d rows" !count;

There are two major benefits to using ~filter in this way:

  1. we don't waste any time parsing the rows since ~filter operates on raw rows and we're dropping them all by always returning false

  2. we're guaranteed to run in constant memory, no matter the XLSX's internal layout

Xlsx.parse_date

XLSX dates are stored as floats. This function converts from a float to a Date.t.

Note: the XLSX type Date exists, but it's rarely ever used.

Xlsx.parse_datetime

XLSX datetimes are stored as floats. This function converts from a float to a Time_float.t.

Xlsx.index_of_column

This function converts a column reference such as "D7" or "AA2" to a 0-based column index.

SZXX.Xlsx.index_of_column "D7" (* returns 3 because column D is the 4th column (0-based index) *)
SZXX.Xlsx.index_of_column "AA2" (* returns 26 because column AA is the 27th column (0-based index) *)

Xlsx.Expert (module)

See xlsx.mli.

SZXX.Xml

This XML parser is not fully spec-compliant: it does not attempt to validate schemas or follow references.

But it's simple, efficient and sufficient for the vast majority of use cases.

It can be configured to parse HTML. It can also stream nested elements out of large XML files without having to hold the whole file in memory.

open! Core
open Eio.Std

let parse_xml xml_path =
  let open SZXX in
  let doc =
    Eio.Path.with_open_in
      Eio.Path.(Eio.Stdenv.fs env / xml_path)
      (fun file -> Xml.parse_document (Feed.of_flow file))
    |> Result.ok_or_failwith
  in
  (* Do something with the doc using the utilities in Xml.DOM *)

To operate on the document and its children, use the utility functions in the Xml.DOM module.

This module transparently unescapes text nodes. Your text nodes will contain the correct e.g. Fast & Furious 🏎️ instead of Fast &amp; Furious &#x1F3CE;&#xFE0F;.

Xml.parse_document

This function progressively assembles an XML document while reading from the raw input. It begins parsing without having to read the whole input in its entirety.

To extract (stream out) specific nested elements instead of fully parsing (very) large documents, use Xml.stream_matching_elements instead.

Xml.parse_document ?parser ?strict feed
Arguments:
  • parser

    • Override the default parser.

    • Make your own parser with Xml.SAX.make_parser or pass Xml.html_parser.

  • strict

    • Default: true.

    • When false, unclosed elements are treated as self-closing elements, HTML-style. For example a <br> without a matching </br> will be treated as a self-closing <br />.

  • feed

    • A producer of raw input data. Create a Feed.t by using the SZXX.Feed module.

Returns:

Xml.document or an error message.

Working with HTML

Let's adapt the previous example and add HTML support:

  • add ~parser:Xml.html_parser

  • add ~strict:false

open! Core
open Eio.Std

let parse_html html_path =
  let open SZXX in
  let doc =
    Eio.Path.with_open_in
      Eio.Path.(Eio.Stdenv.fs env / html_path)
      (fun file ->
        Xml.parse_document
          ~parser:Xml.html_parser (* for HTML *)
          ~strict:false (* for HTML *)
          (Feed.of_flow file) )
    |> Result.ok_or_failwith
  )
  in

  (* Let's extract the text of the first <h1> directly within the first <div> directly within <body> *)
  let text =
    doc.top |> Xml.DOM.(get [ dot "body"; dot "div" ]) |> Option.bind ~f:(Xml.DOM.dot_text "h1")
  in

Xml.parse_document_from_string

Convenience function equivalent to Xml.parse_document (Feed.of_string some_string).

Xml.stream_matching_elements

This function progressively assembles an XML document, but every element that matches filter_path is passed to on_match instead of being added to the DOM. This "shallow DOM" is then returned. It begins parsing without having to read the whole input in its entirety.

Xml.stream_matching_elements ?parser ?strict ~filter_path ~on_match feed
Arguments:
  • parser

    • Override the default parser.

    • Make your own parser with Xml.SAX.make_parser or pass Xml.html_parser.

  • strict

    • Default: true.

    • When false, unclosed elements are treated as self-closing elements, HTML-style. For example a <br> without a matching </br> will be treated as a self-closing <br />.

  • feed

    • A producer of raw input data. Create a Feed.t by using the SZXX.Feed module.

  • filter_path

    • Indicates which part of the DOM should be streamed out instead of being stored in the DOM.

    • For example ["html"; "body"; "div"; "div"; "p"] will emit all the <p> tags nested inside exactly 2 levels of <div> tags in an HTML document.

  • on_match

    • Called on every element that matched filter_path.

Returns:

A "shallow" Xml.document or an error message.

Xml.DOM (module)

This module contains various utilities to traverse and process XML documents. See xml.mli.

  • Xml.DOM.get_attr: get attribute value

  • Xml.DOM.dot: get first child element by tag name

  • Xml.DOM.dot_text: get first child element's inner text by tag name

  • Xml.DOM.at: get the nth child element

  • Xml.DOM.filter_map: filter_map child elements

  • and more, they're all documented in xml.mli

Xml.Expert (module)

See xml.mli.

SZXX.Zip

This ZIP parser supports every "normal" type of ZIP found in the wild: every type and subtype of ZIP 2.0 and 4.5, with compression methods 0 or 8.

Other types of ZIPs are rare and, in practice, not realistically expected to be understood by applications other than the one that created it.

There are two ways to use SZXX.Zip:

  • if you only need a subset of files stored in a ZIP AND your ZIP is a file (not a stream)

    • ➡️ use Zip.index_entries and Zip.extract_from_index

  • otherwise:

    • ➡️ use Zip.stream_files

Zip.stream_files

This function processes a ZIP one file at a time.

For each file, it invokes a callback to ask what it should do with the file, then processes that file, adds the outcome to the output Sequence, then invokes callback for the next file and so on.

In this example, we will extract all files that end in .jpg, skipping over any other files contained within the ZIP archive.

open! Core
open Eio.Std

let save_jpg ~filename contents = ...
  (* Let's pretend this function saves a .jpg file somewhere *)

let unzip_and_save_jpgs zip_path =
  let open SZXX in
  (* The Switch receives any parsing errors *)
  Switch.run @@ fun sw ->

  let file = Eio.Path.(open_in ~sw (Eio.Stdenv.fs env / zip_path)) in

  let callback = function
    | ({ filename; _ } : Zip.entry) when String.is_suffix (String.lowercase filename) ~suffix:".jpg" ->
      (* Here we'll simply extract each .jpg file into a string *)
      Zip.Action.String
    | _ ->
      (* All other files: skip/ignore them *)
      Zip.Action.Fast_skip
  in

  let seq = Zip.stream_files ~sw ~feed:(Feed.of_flow file) callback in

  Sequence.iter seq ~f:(fun (entry, data) ->
    match data with
    | Zip.Data.String contents -> save_jpg ~filename:entry.filename contents
    | _ -> () )

Note that in a real world scenario, we should process each JPG in small chunks instead of decompressing each of them into a (potentially large) string. For that we would use Zip.Action.Fold_string or Zip.Action.Fold_bigstring.

SZXX will call callback for each file it encounters within the ZIP archive. You must choose an Action for SZXX to perform on each file.

Actions:
  • Action.Skip

    • Skip over the compressed bytes of this file without attempting to decompress them.

    • It will still validate the file's integrity as usual

  • Action.Fast_skip

    • Skip over the compressed bytes without attempting to decompress or validate them

  • Action.String

    • Collect the whole decompressed file into a single string.

  • Action.Bigstring

    • Collect the whole decompressed file into a single bigstring.

    • More efficient than Action.String if you don't need to convert the final result into a string later on.

  • Action.Fold_string f

    • Fold this file into a final state using function f, in string chunks of ~8192 bytes.

  • Action.Fold_bigstring f

    • Fold this file into a final state using function f, in bigstring chunks of ~8192 bytes.

    • IMPORTANT: this Bigstring.t is volatile! It's only safe to read from it until the end of function f (the "folder" function). If you need to access the data again later, make a copy of it before the end of function f.

  • Action.Parse p

    • Progressively apply the Angstrom.t parser p to the file while it is being decompressed.

    • Action.Parse expects p to consume all bytes and leave no trailing junk bytes after a successful parse.

  • Action.Parse_many { parser; on_parse }

    • Repeatedly apply the Angstrom.t parser to the file while it is being decompressed.

    • It invokes on_parse on each parsed value.

    • Action.Parse_many expects the file to end on a complete parse and leave no trailing junk bytes.

  • Action.Terminate

    • Abruptly terminate processing of the ZIP archive.

    • The output Sequence will finish with a Data.Terminate element.

    • SZXX stops reading from the Feed.t immediately.

Zip.stream_files ~sw ~feed callback
Arguments:
  • sw

    • A regular Eio.Switch.t. This is where parsing errors go.

  • feed

    • A producer of raw input data. Create a Feed.t by using the SZXX.Feed module.

  • callback

    • A function called on every file found within the ZIP archive.

    • You must choose an Action (Zip.Action.t) to perform over each file encountered within the ZIP archive.

Returns:

A Sequence.t of all files within the archive. The order of the files passed to the callback and on the Sequence matches the arrangement of the files within the ZIP.

The Sequence contains Zip.Data.t elements that match the Zip.Action.t returned by the callback: Zip.Action.Fold_string yields a Zip.Data.Fold_string on the Sequence, Zip.Action.Skip yields Zip.Data.Skip, etc.

Note that a Sequence.t is an ephemeral data structure. Each element can only be seen once, so for example calling Sequence.is_empty will attempt to generate the next element in the Sequence to return true if there was one, but that element will be lost.

SZXX generates elements on demand (lazily), meaning that it will not begin parsing until you begin reading from the Sequence.

Note that you must consume the Sequence from within the Switch. Attempting to read from the Sequence from outside the Switch will fail with Invalid_argument "Coroutine has already failed". The Sequence's contents can safely leave the Switch.

Zip.index_entries

This function finds and returns the ZIP's table of contents (Zip.entry list).

You can then extract files one by one (in any order) using Zip.extract_from_index.

Zip.index_entries file
Arguments:
  • file

    • A file opened with Eio.Path.open_in or Eio.Path.with_open_in.

Returns:

Zip.entry list

Zip.extract_from_index

This function extracts a single file from a ZIP file.

open! Core
open SZXX

let get_file_contents zip_path =
  Eio.Path.(with_open_in (Eio.Stdenv.fs env / zip_path)) (fun file ->
    let entries = Zip.index_entries file in
    let entry = List.filter entries ~f:(fun entry -> ...) in

    let file_contents =
      match Zip.extract_from_index file entry Zip.Action.String with
      | Zip.Data.String s -> s
      | _ -> assert false
    in

    file_contents
  )
Zip.extract_from_index file entry action

See Actions for the meaning of Zip.Action.* and Zip.Data.*

Arguments:
  • file

    • A file opened with Eio.Path.open_in or Eio.Path.with_open_in.

  • entry

    • A Zip.entry value returned by Zip.index_entries

  • action

Returns:

A Zip.Data.t value that matches your action argument.

FAQ

Why am I getting a "Coroutine" error?

Because you are reading a Sequence from outside the Switch that populates the Sequence.

This restriction applies to the following 3 SZXX functions that take a Switch argument and return a Sequence:

  • SZXX.Xlsx.stream_rows_double_pass

  • SZXX.Xlsx.stream_rows_single_pass

  • SZXX.Zip.stream_files

To fix your code: only interact with the Sequence while inside the Switch. In other words: do not let the Sequence escape the Switch scope.

The Sequence's contents can safely leave the Switch.

Is it fast?

Given similar optimization efforts, streaming parsers (like SZXX) should always be slower than parsers that deserialize a whole file into memory and read from that.

However, SZXX has received extensive performance optimizations, to the point that SZXX is faster than many non-streaming XLSX libraries.

It takes a lot of CPU work to extract each row from an XLSX file:

  • the ZIP format was designed for floppy disks and old hard drives, not parallelism or SSDs

  • the XML format is quite verbose and inefficient

  • the XLSX format requires reading and parsing a lot of XML data just to produce a single row of XLSX output

All in all:

  • SZXX.Zip and SZXX.Xml are fast

  • SZXX.Xlsx is extremely fast for a streaming parser, and comparable in performance to the average non-streaming parser that loads everything into memory.

Using 1 core on an ancient 2015 Macbook Pro, SZXX processes a gigantic 28-column x 1,048,576-row XLSX file in 55 seconds using only 9MB of memory. In 2022, the same file would take 70 seconds to open in LibreOffice using 2 cores and 1.8GB of memory, but improvements to LibreOffice have recently brought it down to 36 seconds and 1.1GB.

Where SZXX cannot be beaten is memory usage and predictability.

Any performance tips?

Use flambda. Install the OPAM package named ocaml-option-flambda and add the following to your dune (executable):

(ocamlopt_flags -O3)

Does it work on Windows?

Maybe. Probably. It depends on Eio's level of Windows support. You'll probably need OCaml 5.1 too.

Does it work in the browser?

Version 3 worked but needed a special fork of angstrom.

Version 4 will work too once Eio's level of JS support is sufficient. It may very well be sufficient today already.

You will definitely need to edit your dune files' (libraries) section.

Replace

(libraries SZXX some-other-lib)

with

(libraries checkseum.ocaml SZXX some-other-lib)

checkseum.ocaml has to be before SZXX for it to work.

You may or may not also need to use the aforementioned angstrom fork.

Please give it a try and report back.

Dependencies (6)

  1. eio_main >= "0.12"
  2. decompress >= "1.4.1"
  3. core >= "v0.16.0"
  4. angstrom >= "0.15.0"
  5. dune >= "1.9.0"
  6. ocaml >= "5.0.0"

Dev Dependencies (3)

  1. ppx_deriving_yojson >= "3.5.2" & with-test
  2. yojson with-test
  3. alcotest with-test

Used by

None

Conflicts

None

OCaml

Innovation. Community. Security.