This seems like it would make the wikipedia search result worse, but it drastically improves the result quality!
This is because wikipedia has a lot of articles that each talk about a lot of irrelevant concepts, and indexing the entire document means tangentially relevant results tend to displace the most relevant results.
Modify processingiterator to be constructed via a factory, to enable re-use of its backing executor service.
This reduces thread churn in the converter sideloader style processing of regular crawl data.
Route the sizeHint from the input parquet file to SideloadProcessing, so that it can set sideloadSizeAdvice appropriately, instead of using a fixed "large" number.
This is necessary to populate the KNOWN_URL column in the domain data table, which is important as it is used in e.g. calculating how far to re-crawl the site in the future.
Use ProcessingIterator to fan out processing of documents across more cores, instead of doing all of it in the writer thread blocking everything else with slow single-threaded processing.
This commit adds a safety check that the URL of the document is from the correct domain.
It also adds a sizeHint() method to SerializableCrawlDataStream which *may* provide an indication if the stream is very large and benefits from sideload-style processing (which is slow).
It furthermore addresses a bug where the ProcessedDomain.write() invoked the wrong method on ConverterBatchWriter and only wrote the domain metadata, not the rest...
The processor normally retains the domain data in memory after processing to be able to do additional site-wide analysis. This works well, except there are a number of outlier websites that have an absurd number of documents that can rapidly fill up the heap of the process.
These websites now receive a simplified treatment. This is executed in the converter batch writer thread. This is slower, but the documents will not be persisted in memory.
With the new crawler modifications, the crawl data comes in a slightly different order, and a result of this is that we can optimize the converter. This is a breaking change that will be incompatible with the old style of crawl data, hence it will linger as a branch for a while.
The first step is to move stuff out of the domain processor into the document processor.
Since the sideloaders don't populate the documents list in ProcessedDomain to keep the memory footprint manageable, the code that estimates knownUrls etc. will set them to zero, which has negative effects on their ranking. This change will populate them with a bullshit value within a sane ballpark, ensuring that these domains show up in the rankings.
Make some temporary modifications to the CrawledDocument model to support both a "big string" style headers field like in the old formats, and explicit fields as in the new formats. This is a bit awkward to deal with, but it's a necessity until we migrate off the old formats entirely.
The commit also adds a few tests to this logic.
The size of the ArrayBlockingQueue in ConverterWriter.java has been reduced from 4 to 1. This change aims to reduce the memory utilization by not having fully processed domains piling up in RAM. This may cause the writer to go idle in waiting for new data, but that may be preferable to an OOM.
We do both ip2location and ASN data.
The change also adds some keywords based on autonomous system information, on a somewhat experimental basis. It would be neat to be able to e.g. exclude cloud services or just e.g. cloudflare from the search results.
This variable had a very confusing name, and was dangerously easy to use in the wrong place with the result of getting something that only works as expected half the time.
Ideally this class needs an overhaul, the assumptions it makes about domain names aren't great.
This variable had a very confusing name, and was dangerously easy to use in the wrong place with the result of getting something that only works as expected half the time.
Ideally this class needs an overhaul, the assumptions it makes about domain names aren't great.
In encyclopedia, add a class "mw-content-text" that the WikiSpecialization class is looking for during pruning to give the articles a more fair treatment.
Also add generator keywords based on the generator type provided, to ensure that these documents show up in appropriate filters.
Further, add a new document flag value 'Sideloaded' to be able to distinguish these entries.
There really is no fantastic place to put this logic, but we need to remove entries with an X-Robots-Tags header where that header indicates it doesn't want to be crawled by Marginalia.
We want to mute some of these records so that they don't produce documents, but in some cases we want a document to be produced for accounting purposes.
Added improved tests that reach for known resources on www.marginalia.nu to test the behavior when encountering bad content type and 404s.
The commit also adds some safety try-catch:es around the charset handling, as it may sometimes explode when fed incorrect data, and we do be guessing...
This commit updates CrawlingThenConvertingIntegrationTest with additional tests for invalid, redirecting, and blocked domains. Improvements have also been made to filter out irrelevant entries in ParquetSerializableCrawlDataStream.
This update includes the addition of timestamps to the parquet format for crawl data, as extracted from the Warc stream.
The parquet format stores the timestamp as a 64 bit long, seconds since unix epoch, without a logical type. This is to avoid having to do format conversions when writing and reading the data.
This parquet field populates the timestamp field in CrawledDocument.
Add an optional new field to CrawledDocument containing information about whether the domain has cookies. This was previously on the CrawledDomain object, but since the WarcFormat requires us to write a WarcInfo object at the start of a crawl rather than at the end, this information is unobtainable when creating the CrawledDomain object.
Also fix a bug in the deduplication logic in the DomainProcessor class that caused a test to break.
This commit further cleans up the warc->parquet conversion. It fixes issues with redirect handling in WarcRecorder, adds support information about redirects and errors due to probe failure.
It also refactors the fetch result, body extraction and content type abstractions.
This commit is in a pretty rough state. It refactors the crawler fairly significantly to offer better separation of concerns. It replaces the zstd compressed json files used to store crawl data with WARC files entirely, and the converter is modified to be able to consume this data. This works, -ish.
There appears to be some bug relating to reading robots.txt, and the X-Robots-Tag header is no longer processed either.
A problem is that the WARC files are a bit too large. It will probably be likely to introduce a new format to store the crawl data long term, something like parquet; and use WARCs for intermediate storage to enable the crawler to be restarted without needing a recrawl.
At this stage, the crawler will use the WARCs to resume a crawl if it terminates incorrectly.
This is a WIP commit, since the warc files are not fully incorporated into the work flow, they are deleted after the domain is crawled.
The commit also includes fairly invasive refactoring of the crawler classes, to accomplish better separation of concerns.
This is the same as the prefix for the IP address, but I don't think that substantially matters, the as two have such different namespaces there can be no confusion.
In this commit, GeoIP-related classes are refactored and relocated to a common library as they are shared across multiple services.
The crawler is refactored to enable the GeoIpBlocklist to use the new GeoIpDictionary as the base of its decisions.
The converter is modified ot query this data to add a geoip:-keyword to documents to permit limiting a search to the country of the hosting server.
The commit also adds due BY-SA attribution in the search engine footer for the source of the IP geolocation data.
Refactored the getDocumentsStream method in EncyclopediaMarginaliaNuSideloader to use the newly extracted ProcessingIterator class that encapsulates processing a stream of results from e.g a database query in parallel and returning the computed results as an iterator.
The iterator was also improved on to be more reliable, previous versions of the logic would sometimes deadlock due to false positives in hasMore().
The commit updates EncyclopediaMarginaliaNuSideloader to include the AnchorTextKeywords in processing documents, aiding search result relevance.
It also removes old test-related functionality and a large but fairly useless test previously used to debug a specific problem, to the detriment of the overall code quality.
The converter was not properly initiating the external links for each domain, causing an NPE in conversion. This needs to be loaded later since we don't know the domain we're processing until we've seen it in the crawl data.
Also made some refactorings to make finding converter bugs easier, and finding the related domain less awkward from the SerializableCrawlData interface.
The code now includes an additional function in the DomainProcessor class that checks if a domain is associated with academia. An academic domain is identified by the ".edu" TLD, or fits a specific regex pattern matching domains like *.ac.ccTld or *.edu.ccTld.
If these conditions are met, the search term "special:academia" is added to the domain.
The existing academia search filter uses personalized pagerank to select academia-adjacent domains, but it isn't working very well. The hope is that filtering on domain names will be more effective, and that it can supplant the ranking-based approach.
Partially hook in the WarcRecorder into the crawler process. So far it's not read, but should record the crawled documents.
The WarcRecorder and HttpFetcher classes were also refactored and broken apart to be easier to reason about.
This update includes the integration of the jwarc library and implements support for Warc file sideloading, as a first trial integration with this library.
