fs2-pdf is a Scala library for manipulating PDF files in fs2 streams using scodec for parsing.

Prevalent PDF manipulation tools like iText require the whole file to be read into memory, making it hard to estimate the memory footprint due to large images and imposing a hard boundary for the document file size of Int.MaxValue.

"com.springernature" %% "fs2-pdf" %  "0.1.0-RC8"

The provided fs2 pipes convert pdf data in the shape of a Stream[IO, Byte] into data type encodings and back to byte streams. Raw data is processed with scodec and stored as BitVectors and ByteVectors.

import fs2.pdf._

The coarsest useful data types are provided by wm.pdf.PdfStream.topLevel, producing the ADT TopLevel:

case class IndirectObj(obj: pdf.IndirectObj)
case class Version(version: pdf.Version)
case class Comment(data: pdf.Comment)
case class Xref(version: pdf.Xref)
case class StartXref(startxref: pdf.StartXref)

PDFs consist of a series of objects that look like this:

% a compressed content stream object
3 0 obj
<</Filter /FlateDecode /Length 19>>
stream
<binary data...>
endstream
endobj

% an array object
112 0 obj
[/Name 5 4 0 R (string)]
endobj

% a dict object
2624 0 obj
<</Count 48 /Kids [2625 0 R 2626 0 R 2627 0 R 2628 0 R 2629 0 R 2630 0 R] /Parent 2623 0 R /Type /Pages>>
endobj

These are encoded as IndirectObj, with the stream being optional. There are special object streams, in which a stream contains more objects; those are decoded in a later stage.

At the very beginning of a document, the version header should appear:

%PDF-1.7
%âã��

The second line is optional and indicates that the PDF contains binary data streams.

At the end of a document, the cross reference table, or xref, indicates the byte offsets of the contained objects, looking like this:

xref
0 1724
0000000000 65535 f 
0000111287 00000 n 
0000111518 00000 n 
0000111722 00000 n 
0000111822 00000 n 
0000112053 00000 n 
...
0000111175 00000 n 
trailer
<<
/ID [<9154668ac56ee69570067970a0db0b0a> <3ddbb5faba07f5306b8feb50afd4225c> ]
/Root 1685 0 R
/Size 1724
/Info 1683 0 R
>>
startxref
1493726
%%EOF

The dictionary after the trailer keyword contains metadata, in particular the /Root reference pointing to the object describing the pages.

The number after the startxref keyword denotes the byte offset of the xref keyword for quicker seeking in viewer apps.

Multiple xrefs may occur in a document under two conditions:

• linearized PDFs, where an additional xref at the beginning of the document references only the first page, for optimized loading
• incrementally updated PDFs, allowing authoring tools to append arbitrarily many additional objects with xrefs

Xrefs can be compressed into binary streams. In that case, only the part starting with startxref will be at the end of the file, and the StartXref variant will encode this part.

In order to use this initial encoding, pipe a stream of bytes through PdfStream.topLevel:

val raw: Stream[IO, Byte] =
  openAPdfFile

val topLevel: Stream[IO, TopLevel] =
  raw.through(PdfStream.topLevel)

For a slightly more abstract encoding, TopLevel can be transformed into Decoded with the variants:

case class DataObj(obj: Obj)
case class ContentObj(obj: Obj, rawStream: BitVector, stream: Uncompressed)
case class Meta(xrefs: NonEmptyList[Xref], trailer: Trailer, version: Version)

Here, the distinction between objects with and without streams is made and the metadata is aggregated into a unique record containing all xrefs, the aggregated trailer and the version.

To use this:

val decoded: Stream[IO, Decoded] =
  raw.through(PdfStream.decode(Log.noop)) // you can provide a real logger with `fs2.pdf.Log.io`

For another level of abstraction, the Element algebra represents semantics of objects:

object DataKind
{
  case object General
  case class Page(page: pdf.Page)
  case class Pages(pages: pdf.Pages)
  case class Array(data: Prim.Array)
  case class FontResource(res: pdf.FontResource)
}

case class Data(obj: Obj, kind: DataKind)

object ContentKind
{
  case object General
  case class Image(image: pdf.Image)
}

case class Content(obj: Obj, rawStream: BitVector, stream: Uncompressed, kind: ContentKind)

case class Meta(trailer: Trailer, version: Version)

To use this:

val elements: Stream[IO, Element] =
  raw.through(PdfStream.elements(Log.noop))

A stream of indirect objects can be encoded into a pdf document, with automatic generation of the cross reference table.

val reencoded: Stream[IO, Byte] =
  decoded
    .through(Decoded.parts)
    .through(WritePdf.parts)
    .through(Write.bytes("/path/to/file.pdf"))

or

val reencoded: Stream[IO, Byte] =
  elements
    .through(Element.parts)
    .through(WritePdf.parts)
    .through(Write.bytes("/path/to/file.pdf"))

The intermediate data type Part is used to carry over the trailer into the encoder. Instead of Decoded.parts and WritePdf.parts, you could also use Decoded.objects and WritePdf.objects, but then you would have to specify a trailer dictionary for the encoder. Decoded.parts extracts this information from the input trailer.

Since you won't just want to reencode the original data, a step in between decoding and encoding should manipulate it. The pipes in Rewrite are used in Decoded.parts, and they allow more complex transformations by keeping a state when analyzing objects and using it to create additions to the document.

The rewrite works in two stages, collect and update:

case class PagesState(pages: List[Pages])
val initialState: PagesState = PagesState(Nil)
val result: Stream[IO, ByteVector] =
  elements
    .through(Rewrite(initialState)(collect)(update))

In collect, every Element (or Decoded) will be evaluated by a stateful function that allows you to prevent objects from being written and instead collect them for an update:

def collect(state: RewriteState[PagesState]): Element => Pull[IO, Part[Trailer], RewriteState[PagesState]] = {
  case Element.Data(_, Element.DataKind.Pages(pages)) =>
    Pull.pure(state.copy(state = state.state.copy(pages = pages :: state.state.pages)))
  case Element.obj(obj) =>
    Pull.output1(Part.Obj(obj)) >> Pull.pure(state)
  case Element.Meta(trailer, _) =>
    Pull.pure(state.copy(trailer = Some(trailer)))
}

Here we get our state, wrapped in RewriteState, which tracks the trailer, and an Element passed into our function. The output is fs2.Pull, on which you can call Pull.output1 to instruct the stream to emit a PDF part, and Pull.pure to return the updated state.

There is a variant Rewrite.simple that hides the Pull from these signatures and instead expects you to return (List[Part[Trailer]], RewriteState[PagesState]).

In this example, we match on DataKind.Pages and do not call Pull.output1 in this case, but add the pages to our state. In the case of any other object, which we match with the convenience extractor Element.obj, we just pass through as a Part.Obj. Finally, the trailer has to be carried over in the state.

In update, we use the collected pages to do some analysis and then write them back to the stream:

val fontObj(number: Long): IndirectObj =
  ???

def update(update: RewriteUpdate[PagesState]): Pull[IO, Part[Trailer], Unit] =
  Pull.output1(Part.Trailer(update.trailer)) >> update.state.pages.traverse_ {
    case Pages(index, data, _, true) =>
      val updatedData = data ++ Prim.dict("Resources" -> Prim.dict("Font" -> Prim.dict("F1" -> Prim.Ref(1000L, 0))))
      Pull.output1(fontObj(1000L)) >>
      Pull.output1(Part.Obj(IndirectObj(Obj(index, updatedData), None)))
    case Pages(index, data, _, false) =>
      Pull.output1(Part.Obj(IndirectObj(Obj(index, data), None)))
  }

RewriteUpdate is the same as RewriteState, except that the trailer isn't optional anymore. If there is no trailer in the state at the end of the stream, an error is raised.

In this function, we first emit the trailer, then we iterate over our collected pages and match on the boolean root field. If we found the root page tree object, we first emit a custom font descriptor (not implemented here), then add a reference to it to the page root's Resource dictionary (this is of course an incomplete simplification). For all other pages objects, we just write the original data.

val result: IO[ValidatedNel[String, Unit]] = raw.through(PdfStream.validate(Log.noop))

Linearization is not possible at the moment, since the linearization parameter dict is the first object and needs information that is only available later, like the total file size.

A heuristical method for keeping already linearized documents intact is in development.

ops/sbt test

If you work at SpringerNature and have access to this project's pipeline, you can trigger a deployment to Maven with the script at ops/trigger-publish.bash, which will verify some conditions and start the publish pipeline job.

Copyright 2020 SpringerNature

fs2-pdf is licensed under the Apache License 2.0