* [[:en:Barack Obama]] can now be expanded in 77 seconds using 330MB RAM,
while it would prevously run out of RAM after ~30 minutes. Wohoooo!
The token transform framework rework really paid off.
* 303 parser tests are passing in the new record time of 5.5 seconds. Two more
tests are passing since these tests expect the day of the week to be
Thursday. Won't be the case tomorrow.
Change-Id: I56e850838476b546df10c6a239c8c9e29a1a3136
* All parser pipelines including tokenizer and DOM stuff are now constructed
from a 'recipe' data structure in a ParserPipelineFactory.
* All sub-pipelines of these can now be cached
* Event registrations to a pipeline are directly forwarded to the last
pipeline member to save relatively expensive event forwarding.
* Some APIs for on-demand expansion / format conversion of parameters from
parser functions are added:
param.to('tokens/expanded', cb)
param.to('text/wiki', cb) (this does not work yet)
All parameters are additionally wrapped into a Param object that provides
method for positional parameter naming (.named() or conversion to a dict
(.dict()).
* The async token transform manager is now separated from a frame object, with
the frame holding arguments, an on-demand expansion method and loop checks.
* Only keys of template parameters are now expanded. Parser functions or
template arguments trigger an expansion on-demand. This (unsurprisingly)
makes a big performance difference with typical switch-heavy template
systems.
* Return values from async transforms are no longer used in favor of plain
callbacks. This saves the complication of having to maintain two code paths.
A trick in transformTokens still avoids the construction of unneeded
TokenAccumulators.
* The results of template expansions are no longer buffered.
* 301 parser tests are passing
Known issues:
* Cosmetic cleanup remains to do
* Some parser functions do not support async expansions yet, and need to be
modified.
Change-Id: I1a7690baffbe8141cadf67270904a1b2e1df879a
Renamed Selection method to more suitable name.
Misc cleanup
Patchset 2, whitespace cleanup
Patchset 3: Change values used with selection direction to -1 or 1
1 for left to right (normal)
-1 for right to left (opposite)
Change-Id: If9ecc721ace1c7550903170f92395947f1ccc22c
It's not being used at all, and it's broken because
this.lengthDifference is set to zero regardless of what length
difference the operations passed into the constructor might cause
Change-Id: I3b7a312a1920347e7bf34df88a05bf6f2ff11f7d
* Changed splice to check all elements about to be inserted are allowed before inserting any of them so that catching an exception leaves you in a sane state
* Fixed the order of execution of parent class constructors in ve.dm.LeafNode and ve.dm.TwigNode so that canHaveChildren and canHaveGrandchildren produce correct values and added tests to ensure these methods are correctly inherited in subclasses
* Added tests that check for exceptions when adding nodes that can have children to nodes that can not have grandchildren
* Added test that check for events being emitted before and after splicing, including that beforeSplice should be emitted even in cases where a splice fails and throws an exception because the nodes are incompatible (but afterSplice is not called in this case) since beforeSplice might modify the nodes in some way before the compatibility tests are run
Change-Id: Id12aea995a42c26ff63a74ae3d31f2bf455759e3
* Moved getParent and getRoot from ve.dm.Node back to ve.Node
* Fixed use of getElementLength that should have been changed to getOuterLength, but was changed to getLength (oops)
Change-Id: Ibe5b855aef533dcd493f762a8a02c6a11ce6e7de
In this commit several methods (child node add/remove and parent/root modification) were also moved to ve.dm.BranchNode ve.dm.Node respectively. ve.Node and ve.BranchNode are immutable. ve.dm.Node and ve.dm.BranchNode are mutable. Other subclasses of ve.Node and ve.BranchNode should implement functionality to mimic changes made to a data model.
Change-Id: Ia9ff78764f8f50f99fc8f9f9593657c0a0bf287e
Ground-up rewrite of the data model. Putting this in the ve2 directory for now so we still have the old code floating around.
Main changes so far in this rewrite:
* Renamed hasChildren() to canHaveChildren()
* Added canHaveGrandchildren()
* Added a new node type TwigNode that can have children but not grandchildren (so all of its children must be LeafNodes)
* Implemented push/pop/shift/unshift as wrappers around splice()
* Renamed getElementType() to getType(). Nodes now take a string as a type, and the element stuff is gone and won't be back
* Removed clearRoot(), replaced it with setRoot( this ) where needed
Change-Id: I23f3bb1b4a2473575e5446e87fdf17af107bacf6
This is a bit better than cloning tokens wholesale, but not by much. There is
a lot of potential for much better per-token caching with reduced token
cloning. Need to map out all dependencies besides token attributes expanded
from template parameters or other scoped state. Even if tokens themselves
don't need transformation, they might still need to be considered for other
token transformers, so simply keeping the final rank won't quite work even if
the token itself is fully transformed. As a minimum, a shallow clone would
need to be made and the rank reset (as in env.cloneTokens).
Change-Id: I4329113bb21750bae9a635229ed1b08da75dc614
* Added an LRU cache (using the lru-cache node module) for tokenizer output
* Mutation of nested attributes now replaces the containers. A shallow copy of
tokens is sufficient to isolate token transformations. Need to investigate
if we can actually get away without isolation and re-transformation for most
ordinary tokens.
Change-Id: I9136b1d7a1fbcc538183a319d4ecaa290d616fdf
* Ignore safesubst for now
* Remove an unneeded whitelist entry
* Make sure the caption is not lost for thumbs (fix to last commit) and remove
debug print
Change-Id: I243584ed0838cf7c3b4110fe9cdf869272477312
The HTML5 parser we are using to normalize expected HTML output in parserTests
reverses the order of attributes (see
https://github.com/aredridel/html5/pull/53 for the fix). Remove whitelist
entries concerned with this and use the proper order in external image
attributes.
Change-Id: If1868cae05396a150757c85a20473ab756cbcd97
* less verbose logging in noinclude processing and template expansion
* Give priority to the processing of templates transcluded from transclusions
to get closer to depth-first processing. This serves to minimize memory
usage from queued-up tokens.
* Increase the maximum outstanding requests per template retrieval. 10000
amazingly proved too low a limit on some big pages.
* Only process a single template request callback at a time for now
* Add a debug print in the treebuilder wrapper
* Don't treat multiple comments on a single line as a single comment to match
the PHP parser's behavior
Change-Id: I9a86b6d7bec3b9e1f17415daf1bf74170240721a
This has some TODOs still but I want to land it now anyway, and fix the
TODOs later.
* Add this.offsetMap which maps each linear model offset to a model tree node
* Refactor createNodesFromData()
** Rename it to buildSubtreeFromData()
** Have it build an offset map as well as a node subtree
** Have it set the root on the fake root node so that when the subtree
is attached to the main tree later, we don't get a rippling root
update all the way down
** Normalize the way the loop processes content, that way adding offsets
for content is easier
* Add rebuildNodes() which uses buildSubtreeFromData() to rebuild stuff
* Use rebuildNodes() in DocumentSynchronizer
* Use pushRebuild() in TransactionProcessor
* Optimize setRoot() for the case where the root is already set correctly
Change-Id: I8b827d0823c969e671615ddd06e5f1bd70e9d54c
In experiments this dropped the memory consumption further, and reduces the
queuing overhead in the node reactor.
Change-Id: I9409b6ca863b43b7557663bbec9572365059c078
Only call back a few callbacks per reactor iteration from the template fetch
request queue. This changes the expansion pattern from a (memory intensive)
breadth-first expansion to something quite close to depth-first expansion.
Additionally, retrieved pages are quickly added to the page cache so that a
lot of request queuing is avoided in favor of synchronous expansion from the
cache. On pages like Barack Obama that previously ran out of memory after
consuming node's 1.6G heap limit, expansion now runs in relatively constant
100-300M resident (so far, still running).
Change-Id: Ie34a1eeff00d868416de45ef8d289898258f560c
