[
https://issues.apache.org/jira/browse/TIKA-3890?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17620633#comment-17620633
]
Nick Burch commented on TIKA-3890:
----------------------------------
DOCX files are compressed XML. Text compresses very well. Already compressed
images, audio, video don't.
An 8mb word document of pure text could fairly easily produce a 10x that in
text. An 8mb word document that's mostly images could produce just a few bytes
of text
DOCX-specific, you could open the file in POI (use a File to save memory), and
check the size of the word XML stream and the size of any attachments, that'd
give you a vague idea. However, it won't give you a complete answer as the word
XML could have loads of complex stuff in it that doesn't end up with text
output...
Easiest way to know the size of the output is just to parse it on a beefy
machine with suitable restarts / respawning in place, and see what you get!
> Identifying an efficient approach for getting page count prior to running an
> extraction
> ---------------------------------------------------------------------------------------
>
> Key: TIKA-3890
> URL: https://issues.apache.org/jira/browse/TIKA-3890
> Project: Tika
> Issue Type: Improvement
> Components: app
> Affects Versions: 2.5.0
> Environment: OS: OSX, Processor: M1 (ARM), RAM: 64GB, CPU: 8 cores
> Docker container with 5.5GB reserved memory, 6GB limit
> Tika config w/ 2GB reserved memory, 5GB limit
> Reporter: Ethan Wilansky
> Priority: Blocker
>
> Tika is doing a great job with text extraction, until we encounter an Office
> document with an unreasonably large number of pages with extractable text.
> For example a Word document containing thousands of text pages.
> Unfortunately, we don't have an efficient way to determine page count before
> calling the /tika or /rmeta endpoints and either getting back an array
> allocation error or setting byteArrayMaxOverride to a large number to return
> the text or metadata containing the page count. Returning a result other than
> the array allocation error can take significant time.
> For example, this call:
> {{curl -T ./8mb.docx -H "Content-Type:
> application/vnd.openxmlformats-officedocument.wordprocessingml.document"
> [http://localhost:9998/rmeta/ignore]}}
> {quote}{{with the configuration:}}
> {{<?xml version="1.0" encoding="UTF-8" standalone="no"?>}}
> {{<properties>}}
> {{ <parsers>}}
> {{ <parser class="org.apache.tika.parser.DefaultParser">}}
> {{ <parser-exclude
> class="org.apache.tika.parser.ocr.TesseractOCRParser"/>}}
> {{ <parser-exclude
> class="org.apache.tika.parser.microsoft.OfficeParser"/>}}
> {{ </parser>}}
> {{ <parser class="org.apache.tika.parser.microsoft.OfficeParser">}}
> {{ <params>}}
> {{ <param name="byteArrayMaxOverride" type="int">175000000</param>}}
> {{ </params>}}
> {{ </parser>}}
> {{ </parsers>}}
> {{ <server>}}
> {{ <params>}}
> {{ <taskTimeoutMillis>120000</taskTimeoutMillis>}}
> {{ <forkedJvmArgs>}}
> {{ <arg>-Xms2000m</arg>}}
> {{ <arg>-Xmx5000m</arg>}}
> {{ </forkedJvmArgs>}}
> {{ </params>}}
> {{ </server>}}
> {{</properties>}}
> {quote}
> returns: {{"xmpTPg:NPages":"14625"}} in ~53 seconds.
> Yes, I know this is a huge docx file and I don't want to process it. If I
> don't configure {{byteArrayMaxOverride}} I get this exception in just over a
> second:
> {{Tried to allocate an array of length 172,983,026, but the maximum length
> for this record type is 100,000,000.}} which is the preferred result.
> The exception is the preferred result. With that in mind, can you answer
> these questions?
> 1. Will other extractable file types that don't use the OfficeParser also
> throw the same array allocation error for very large text extractions?
> 2. Is there any way to correlate the array length returned to the number of
> lines or pages in the associated file to parse?
> 3. Is there an efficient way to calculate lines or pages of extractable
> content in a file before sending it for extraction? It doesn't appear that
> /rmeta with the /ignore path param significantly improves efficiency over
> calling the /tika endpoint or /rmeta w/out /igmore
> If its useful, I can share the 8MB docx file containing 14k pages.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
