Recent changes to the result ranking mean the no filter mode returns sufficiently good results for most queries that filtering by default just makes the search results more restricted.
* (executor-api) Make executor API talk GRPC
The executor's REST API was very fragile and annoying to work with, lacking even basic type safety. Migrate to use GRPC instead. GRPC is a bit of a pain with how verbose it is, but that is probably a lesser evil. This is a fairly straightforward change, but it's also large so a solid round of testing is needed...
The change set breaks out the GrpcStubPool previously residing in the QueryService, and makes it available to all clients.
ServiceId.name was also renamed to avoid the very dangerous clash with Enum.name().
The boilerplate needed for grpc was also extracted into a common gradle file for inclusion into the appropriate build.gradle-files.
!bang query handling seems to have fallen victim to an overzealous refactoring effort, and broken.
It's now repaired, and a test is in place to ensure we know if it breaks again.
The readme for the array library was extremely out of date. Updating it with accurate information about how the library works, and a demo that should compile.
Also added a system property for disabling the use of sun.misc.Unsafe.
Continues 467ba5be20 by breaking out a constant with the name of the primary ranking set. Also ensures it doesn't get spuriously logged as updated during the secondary updating pass.
This change splits the previous 'repartition' action into two steps, one for recalculating the domain rankings, and one for recalculating the other ranking sets. Since only the first is necessary before the index construction, the rest can be delayed until after...
To avoid issues in handling the shotgun blast of MqNotifications, Service was switched over to use a synchronous message queue instead of an asynchronous one.
The change also modifies the behavior so that only node 1 will push the changes to the EC_DOMAIN database table, to avoid unnecessary db locks and contention with the loader.
Additionally, the change fixes a bug where the index construction code wasn't actually picking up the rankings data.
Since the index construction used to be performed by the index-service, merely saving the data to memory was enough for it to be accessible within the index-construction logic, but since it's been broken out into a separate process, the new process just injected an empty DomainRankings object instead.
To fix this, DomainRankings can now be persisted to disk, and a pre-loaded version of the object is injected into the index-construction process.
The codebase used to have a monkey patched version of gson that made special optimizations for the unusually large JSON files that used to store e.g. crawl data.
Since JSON is no longer used in this fashion, the GSON fork is not needed anymore.
To help distinguish between environments, a system property 'control.appBorder' is added that is injected as a body element border property in the control GUI stylesheets.
IndexJournalWriterPagingImpl was modified to not page on number of entries written, but number of (equivalent uncompressed) bytes written.
Since the failure mode if too much data is written per file is quiet corruption of the index, the former behavior was extremely fragile. The new behavior should consistently ensure that the data is sufficiently small to not cause any integer rollovers.
The change in 6dcc20038c was reverted, as there is really no sane reason to have this configurable in software.
The RandomFileAssembler implementations, introduced in commit 53c575db3f were all acting subtly differently. The RWF implementation wrote BigEndian longs instead of the native endianness used by the other implementations (and expected by the index construction code), further the mmap implementation exposed a bug in LongArray.write() that caused it to create a larger file than necessary.
A test was built to ensure the output of these implementations is equivalent.
To cope with writing large files out of order, the system needs some form of strategy to avoid writing them directly to disk, as this causes insane amounts of disk thrashing. By default, the data is just buffered in RAM. This works well on a large server, but smaller systems struggle.
To help systems with small RAM process large amounts of data, the old RandomWriteFunnel is brought back if the system property 'system.conserve-memory' is set to true. RandomWriteFunnel is buffering the construction by creating a series of small files that pigeonhole the writes into rough neighborhoods, and then it goes over the files one by one to construct one area of the file at a time. This is relatively slow and uses more than twice the disk size.
A new interface RandomFileAssembler is introduced as an abstraction for this operation. A third strategy, direct mmaps, is also introduced if the file is very small (less than 1 GB). In this domain, disk thrashing is unlikely since it will comfortably fit in RAM.
This update enhances the SideloaderProcessing and DocumentClass modules to specially handle sideloaded wiki documents. Wiki content is generally truncated to the first paragraph, which generally tends to be too short to be included independently. An additional DocumentClass (SIDELOAD) has been introduced to suppress the length check in this case.
Wrote a new test to examine the redirect behavior of the crawler, ensuring that the redirect URL is the URL that is reported in the parquet file. This works as intended.
Noticed in the course of this that the crawler doesn't add links from meta-tag redirects to the crawl frontier. Added logic to handle this case, amended the test case to verify the new behavior. Added the meta-redirect case to the HtmlDocumentProcessorPlugin as well, so that we consider it a link between documents in the unlikely case that a meta redirect is to another domain.
The flag is `system.languageDetectionModelVersion`.
* If negative, no model is used.
* If 0, both models are used.
* If 1, the old crappy model is used.
* If 2, the new fasttext model is used.
The flag is `system.languageDetectionModelVersion`.
* If negative, no model is used.
* If 0, both models are used.
* If 1, the old crappy model is used.
* If 2, the new fasttext model is used.
The production configuration assumes all content of interest is 7 bit ASCII, and makes a series of optimizations based on this. This assumption holds poorly in the wild.
Adding an **experimental** system property 'system.noFlattenUnicode', that when set to TRUE, will disable this behavior.
IMPORTANT!! The index needs to be re-constructed when this flag is changed, as different hash functions are selected for the keyword->identifier mappings.
Since the bleed-flags set by the anchor tags logics have been changed to Site and SiteAdjacent, give them a bit of more importance when set together with ExternalLink.
UrlDomain and UrlPath are also only more consistently only rewarded once.
This changeset adds an action for downloading a set of sample data from downloads.marginalia.nu.
It also refactors out some leaky abstractions out of FileStorageService. allocateTemporaryStorage has been renamed allocateStorage. The storage was never temporary in any scenario...
It also doesn't take a storage base, as there was always only one valid option for this input. The allocateStorage method finds the appropriate base itself.
This will avoid having to dig in the message queue to perform this relatively common task.
The control service was also refactored to extract common timestamp formatting logic out of the data objects and into the rendering.
It's a mistake to let it bleed into Title, as this is a high quality signal. We'll co-opt Site and SiteAdjacent instead to reinforce the ExternalLink when count is high.
This avoids concurrent access errors. This is especially important when using Unsafe-based LongArrays, since we have concurrent access to the underlying memory-mapped file. If pull the rug from under the caller by closing the file, we'll get a SIGSEGV. Even with a "safe" MemorySegment, we'll get ugly stacktraces if we close the file while a thread is still accessing it.
So we spin up a thread that sleeps for a minute before actually unmapping the file, allowing any ongoing requests to wrap up. This is 100% a hack, but it lets us get away with doing this without adding locks to the index readers.
Since this is "just" mmapped data, and this operation happens optimistically once a month, it should be safe if the call gets lost.
The sideload instruction in the stackexchange template was updated. The instruction now states that stackexchange data will be loaded from a directory on the server and directs users to a new documentation url for more detailed information.
The method `isApplicable` in the `PlainTextDocumentProcessorPlugin` was refactored to handle a wider range of content types beyond merely "text/plain". It now also handles any content type that starts with "text/plain;", to accomodate contentTypes that append a charset as well.
Removed the need to have to run an external tool to pre-process the data in order to load stackexchange-style data into the search engine.
Removed the tool itself.
This stirred up some issues with the dependencies, that were due to both third-party:ing xz and importing it as a dependency. This has been fixed, and :third-party:xz was removed.
java.lang.Error:s were not handled properly, leading to mismatch in the bookkeeping of the FSMs. These are now caught, acted on, and re-thrown.
MqSynchronousInbox also no longer assumes all exceptions are InterruptedException.
This commit extracts several previously hardcoded configuration properties, and makes then available through system.properties.
The documentation is updated to reflect the change.
Dead code was also removed in the process. CrawlSpecGenerator is left feeling a bit over-engineered still, since it's built for a more general case, where all other implementations but the current one are removed, but we'll leave it like this for now as it's fairly readable still.
Previously, in order to load encyclopedia data into the search engine, it was necessary to use the encyclopedia.marginalia.nu converter to first create a .db-file. This isn't very ergonomic, so parts of that code-base was lifted in as a 3rd party library, and conversion from .zim to .db is now done automatically.
The output file name is based on the original filename, plus a crc32 hash and a .db-ending, to ensure we can recycle the data on repeat loads.
The change adds a new column to the MESSAGE_QUEUE table called AUDIT_RELATED_ID. This field is populated transparently, using a dictionary mapping Thread IDs to Message IDs, populated by the inbox handlers.
The existing RELATED_ID field has too many semantics associated with them,
among other things the FSM code uses them this field in tracking state changes.
The change set also improves the consistency of inbox names. The IndexClient was buggy and populated its outbox with a UUID. This is fixed. All Service2Service outboxes are now prefixed with 'pp:' to make them even easier to differentiate.
It's a confusing default behavior.
This was off for nodes n>1 before as a bandaid since querying indices with no data caused delays and errors. This has been fixed now, so there's no need to do this anymore!
This improves query times, and gets rid of exceptions in the logs when one of the index nodes doesn't have any data loaded, yet is configured to answer queries.
In some scenarios, such as when restoring storage items from json-manifest on db failure, the file storage view would present the items in a non-chronological order. Added a sort() operation to mitigate this.
Use a system.properties file to configure the system. This is loaded statically by MainClass or ProcessMainClass. Update the property names to be more consistent, and update the documentations to reflect the changes.
The EC_DOMAIN_LINK table is deprecated and slated for removal, use QueryClient.getAllDomainLinks() instead.
The ExportDataActor now uses the QueryClient appropriately. The CSV format was also changed to quote the values, to prevent e.g. Excel from interpreting the comma as a decimal separator when previewing the file.
Finally the form for triggering an export was overhauled.
Several tests were manually running migrations in a large copy-paste blob of code. This makes the test less useful as it's possible to break the code while keeping the tests green by introducing a new migration that never gets run in the tests, and it's also difficult to reason about what the tests are doing.
A new test helper library is introduced with a TestMigrationLoader that can both run Flyway migrations, or load specific migrations in the cases a specific set of migrations need to be loaded. Existing tests are migrated to use the new code.
Add a toggle for saving the WARC data generated by the search engine's crawler. Normally this is discarded, but for debugging or archival purposes, retaining it may be of interest.
The warc files are concatenated into larger archives, up to about 1 GB each.
An index is also created containing filenames, domain names, offsets and sizes
to help navigate these larger archives.
The warc data is saved in a directory warc/ under the crawl data storage.
This commit overhauls a lot of the UX for the control service, adding a new actions menu to the nodes views. It has many small tweaks to make the work flow better.
It also adds a new /uploads directory in each index node, from which sideloaded data can be selected. This is a bit of a breaking change, as this directory needs to exist in each index node.
The changeset also makes the control service responsible for flyway migrations. This helps reduce the number of places the database configuration needs to be spread out. These automatic migrations can be disabled with -DdisableFlyway=true.
The commit also adds curl to the docker container, to enable docker health checks and interdependencies.
This adds a docker-compose file 'docker-compose-barebones.yml' which will only start the minimal number of services needed to run a whitelabel Marginalia Search-style search engine, with none of the surrounding frills.
The change also adds a minimal search GUI to the query service, which is also available with JSON results if the appropriate Accept header is provided.
This was caused by incorrect usage of the renderInto() function, which was always buggy and should never be used. This method is removed with this change.
This was caused by incorrect usage of the renderInto() function, which was always buggy and should never be used. This method is removed with this change.
The new converter logic assumes that the crawl data is ordered where the domain record comes first, and then a sequence of document records. This is true for the new parquet format, but not for the old zstd/gson format.
To make the new converter compatible with the old format, a specialized reader is introduced that scans for the domain record before running through the sequence of document records; and presenting them in the new order.
This is slower than just reading the file beginning to end, so in order to retain performance when this ordering isn't necessary, a CompatibilityLevel flag is added to CrawledDomainReader, permitting the caller to decide how compatible the data needs to be.
Down the line when all the old data is purged, this should be removed, as it amounts to technical debt.
The EC_DOMAIN_LINK MariaDB table stores links between domains. This is problematic, as both updating and querying this table is very slow in relation to how small the data is (~10 GB). This slowness is largely caused by the database enforcing ACID guarantees we don't particularly need.
This changeset replaces the EC_DOMAIN_LINK table with a file in each index node containing 32 bit integer pairs corresponding to links between two domains. This file is loaded in memory in each node, and can be queried via the Query Service.
A migration step is needed before this file is created in each node. Until that happens, the actual data is loaded from the EC_DOMAIN_LINK table, but accessed as though it was a file.
The changeset also migrates/renames the links.db file to documents.db to avoid naming confusion between the two.
