[
https://issues.apache.org/jira/browse/PDFBOX-5263?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17425557#comment-17425557
]
Christian Appl commented on PDFBOX-5263:
----------------------------------------
*Notes to self - thinking about whether and how this could be improved*
*Rules:*
- Do not iterate COSStructures - _at all!_ (Iteration costs time, skip steps
wherever possible. Determine all required information exactly in the moment a
change is detected.)
- Everything that is not a COSUpdateInfo is irrelevant and shall be skipped,
not handled, not even be traversed - those must not be touched and must not be
seen.**
- Do not collect information, that is not required whatsoever.
- Do not include objects, that are not being updated! (for signatures that
might be okay, but increments of the size of the original document are
inacceptable.)
- Do not evaluate pathes and reference holders for indirect objects! (Indirect
objects can have multiple parents - processing pathes and reference holders is
costly.)
- Do not cause direct objects to become indirect! (Which somehow conflicts the
previous point, as the parent must be known and updated for direct objects.)
- Collect updated objects in a flat list at a central place, that can be used
by the COSWriter directly, without adapting it further!
Well.... I do not know whether that is a can-do...
*Thoughts:*
- Every object either is direct or indirect. In the end only indirect objects
can be added to an increment, storing direct objects for later evaluation
should be avoided at all costs, as it results in having to find the indirect
object, that we would actually like to update.
- The document already knew all contained indirect objects of a document in 2.0
- possibly that is something we could work with.
- If the parser can tell objects, which changes it causes (hidden
method/state/flag/lock?). Those changes could be ignored, without the object
even having to know it's document context. (However the parser would do that.)
- Dereferencing objects is not a special case, if parsers inform objects about
their activity, as dereferencing is also handled by an ICOSParser, that can do
the same.
- Not all COSUpdateInfo structures are relevant, if a COSUpdateInfo is direct,
it should inform the next higher (containing) indirect object about the update
instead.
- When a COSUpdateInfo is told to be direct, the containing indirect object
could be given to the object - then the object itself could redirect all
updates to the containing structure. (setNeedToBeUpdated)
- Indirect Objects must know a central place to report updates to (reduced core
of COSIncrement) - which is collecting updated indirect objects without further
questions, in an ordered manner. That structure would be creating the increment
by doing so.
- The increment would again be created in realtime, without having to
postprocess the document for the COSWriter, using structures that are as cheap
as can be.
This sounds _somewhat_ plausible and possible....
I will give it another go at the weekend - but I can give no guarantees at all,
that I end up with a thing that actually works.
> Suggestion: Signing actual document changes - Enhancing incremental saving
> --------------------------------------------------------------------------
>
> Key: PDFBOX-5263
> URL: https://issues.apache.org/jira/browse/PDFBOX-5263
> Project: PDFBox
> Issue Type: Improvement
> Components: Parsing, PDModel, Writing
> Affects Versions: 3.0.0 PDFBox
> Reporter: Christian Appl
> Priority: Major
> Fix For: 3.0.0 PDFBox
>
> Attachments: Enhanced_incremental_saving_.patch,
> Enhanced_incremental_saving_PDFBox3.patch, NoSignatureFound.zip,
> Observer_based_incremental_saving(09-13-2021).patch,
> Observer_based_incremental_saving_(09-09-2021-09-09).patch,
> Observer_based_incremental_saving_(12-00-2021-09-08).patch,
> Observer_based_incremental_saving_(15-19-2021-09-08).patch,
> Observer_based_incremental_saving_(17-00-2021-09-07).patch,
> Observer_based_incremental_saving_(fixed_).patch,
> Observer_based_incremental_saving_-_still_eroneous.patch,
> Observer_based_incremental_saving_.patch,
> PDFBOX-5263-YTW2VWJQTDAE67PGJT6GS7QSKW3GNUQR_afterPatch.pdf,
> PDFBOX-5263-YTW2VWJQTDAE67PGJT6GS7QSKW3GNUQR_beforePatch.pdf,
> PDFBOX-5263_Introduce_COSReferenceInfo.patch,
> PDFBOX-5263_Introduce_COSReferenceInfo_(LinkedHashMap).patch,
> PDFBOX-5263_Introduce_state_based_increment_handling_.patch,
> PDFBOX-5263_Introduce_state_based_increment_handling__(POC2).patch,
> PDFBOX-5263_Introduce_state_based_increment_handling__(simplified).patch,
> Prototype_COSChangeObserver_(code_base).patch,
> Prototype__Document_and_reference_holder_aware_COSContext_.patch,
> Updating_context_management_.patch, image-2021-08-23-14-55-24-077.png,
> image-2021-08-26-09-52-33-567.png, image-2021-08-26-09-54-24-897.png,
> image-2021-08-26-10-00-07-383.png, image-2021-08-26-10-02-08-003.png,
> image-2021-08-26-10-03-47-940.png, image-2021-08-26-10-06-42-925.png,
> image-2021-08-26-10-09-12-698.png, image-2021-08-26-10-12-19-265.png,
> image-2021-09-06-17-06-59-667.png, image-2021-09-07-09-35-31-408.png,
> image-2021-09-07-13-33-00-161.png, image-2021-09-07-15-40-59-080.png,
> image-2021-09-08-10-23-44-036.png, image-2021-09-08-11-18-34-211.png,
> image-2021-09-13-14-40-33-049.png, image-2021-09-13-14-41-13-206.png,
> image-2021-09-30-09-33-04-449.png, image-2021-09-30-09-34-19-175.png,
> image-2021-10-07-13-06-40-576.png, image-2021-10-07-13-38-34-817.png,
> out.pdf, out2.pdf
>
>
> *TL;DR:*
> Currently it is rather tedious to create incremental changes in between
> signatures via PDFBox. I attempted to simplify that and wrote a patch.
> This is rather a POC, than an actual suggestion for direct inclusion. (For
> reasons explained later.)
> *Signatures and incremental PDF documents:*
> A typical reason for wanting to sign a document multiple times (creating an
> incremental PDF) is , that in between signatures the document changed and the
> additional signature shall sign the new state of the document.
> If one wanted to implement such incremental changes using PDFBox, he would
> find, that most of the time made changes are completly ignored, when calling
> "saveIncremental".
> As documented for the "saveIncremental" methods and especially the matching
> constructors in "COSWriter", this would require, to identify the "path" of
> all made changes and one would need to set the "needToBeUpdated" flag of all
> elements of that path.
> *But:*
> As documented one would have to have exact understanding of what he did and
> how the PDF standard does implement this, he would have to identify said
> structures and the more complex the changes were, the more tedious this would
> become.
> *Also:*
> Because of the implementation of incremental saving in COSWriter, the whole
> path must be informed that it required an update.
> Resulting in unnecessary large increments, as not all ancestors might
> actually have changed.
> e.g. If one added an image to a preexisting page of the document - the
> contentstream, the resources of the page and the page dictionary would have
> changed. But the "pages" array and all it's ancestors would not have changed
> a bit, but still would have to be informed and included.
> *Assumptions that lead to this patch:*
> - COSWriter should not stop iterating a COSTree just because a parent element
> did not change. It's descendants still could have changed!
> - Externally managing an object´s update state is tedious and error-prone.
> Objects that implement "COSUpdateInfo" should know and manage by themselves
> whether they were freshly created or altered
> (e.g.: A COSDictionary should be able to remember, that a setter had been
> called).
> - If "COSUpdateInfo" objects were self aware and would solve this by
> themselves, it would not be necessary anymore to set update states manually.
> *Problems:*
> The first and obvious problem is, that the initial loading of a document is
> creating and altering new COS structures and we obviously don't want objects
> to observe and remember those changes. An object that is created during
> document initialization must be treated as preexisting.
> However: COSBase is not context aware - it does know it's descendants, but
> neither does it know it's parent, nor does it know it's root.
> If it was, that actually would present the optimal solution, as in that case
> the Object could ask it's root for the current load state and therefore would
> be able to ignore said changes caused by the initial loading of a document.
> But it is not. (My opinion is - it should be! But more on that later.)
> Therefore a a helper named COSUpdateInfoList was implemented, which was
> capable of finding COSUpdateInfo objects in a COS structure, and that allowed
> resetting their update state after loading was completed.
> *Description of the patch:*
> The patch implements selfaware COSUpdateInfo objects, which the COSWriter has
> been adapted to process. PDFBox therefore is capable of monitoring changes in
> realtime and to automatically include altered structures in an incremental
> save of the document, therefore creating increments (or an increment), that a
> signature would sign.
> *Result:*
> Using this patch documents could be created:
> incrementally adding pages, adding contents to pages, adding annotations,
> altering structures, removing structures.
> As far as has been initially tested the resulting documents were valid,
> viewable in a reader and the objects overwritten in increments seemed correct.
> *But -* *Caveat:*
> This patch does introduce atleast one ugly class (most likely you will be
> able to point out more, that could be optimized :)) and that is
> "COSUpdateInfoList" - as already explained: In my opinion such a class should
> not exist, the COSUpdateInfo objects should be context aware and should be
> capable of regulating their own behaviour.
> Whenever the alternatives are to either manage an object externally, or to
> "teach" an object to solve problems autonomously, I will tend to prefer the
> latter... but I did not dare to do that.
> This would require, that either further constructors or setters would have to
> be introduced for COSBase objects, that allowed setting parent/root/context
> for the object.
> Which would result in further massive changes for using applications and
> PDFBox itself - as all instantiations of COSBase objects (PDObjects) would
> have to be adapted.
> However: I would prefer if COSBase objects actually were context aware.
> But as stated... I did not dare to touch it and instead chose the ugly
> workarround, that would introduce yet another iteration over the whole
> COSDocument structure.
> Eliminating COSUpdateInfoList would be preferable!
> *Suggestion:*
> As PDFBox 3 is already changing how documents and objects are handled, I
> would suggest, that also COSBase objects should be made context and selfaware
> in PDFBox 3.
> This would allow simplifying handling COS objects using PDFBox and it would
> allow for an easier and automized handling of incremental saving.
> *Usage example:*
> The following "pseudo code" (actually using simplified Helper classes)
> demonstrates the intended usage:
> !image-2021-08-23-14-55-24-077.png!
> *As always:* Thank you very much for your work and support! I hope this
> suggestion is to your liking.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
