It had been previously assumed that re-writing this function in the style of retain() would make it faster, but it had the opposite effect.
The reason why retain is so fast due to properties of the data that hold true when intersecting document lists, where long runs of adjacent documents are expected, but not when looking up the data associated with the already intersected documents, where the data is more sparse.
IntArray gets the YAGNI axe. The array library had two implementations, one for longs which was used, and one for ints, which only ever saw bit rot. Removing the latter, as all it ever did was clutter up the codebase and add technical debt. If we need int arrays, we fork LongArray again (or add int capabilities to it)
Also cleaning up the interfaces, removing layers of redundant abstractions and adding javadocs.
Finally adding sz=2 specializations to the quick- and insertion sort algorithms. It seems the JIT isn't optimizing these particularly well, this is an attempt to help it out a bit.
Retire search functions that weren't used, including the native implementations. Drop confusing suffixes on search function names. Search functions no longer encode search misses as negative values.
Replaced binary search function with a branchless version that is much faster.
Cleaned up benchmark code.
Retire search functions that weren't used, including the native implementations. Drop confusing suffixes on search function names. Search functions no longer encode search misses as negative values.
Replaced binary search function with a branchless version that is much faster.
Cleaned up benchmark code.
This corrects an annoying bug that had the system crash and burn on first start-up due to a race condition in service initialization, where the services were attempting to access the database before it was properly migrated.
A fix was in principle already in place, but it was running too late and did not prevent attempts to access the as-yet uninitialized database. Move the first boot check into the MainClass instead of the Service constructor.
The change also adds more appropriate docker dependencies to the services to fix rare errors resolving the hostname of the database.
Before the gRPC migration, the system would serve both public and internal requests over HTTP, but distinguish the two using path prefixes and a few HTTP Headers (X-Public, X-Context) added by the reverse proxy to prevent misconfigurations.
Since internal requests meaningfully no longer use HTTP, this convention is just an obstacle now, adding the need to always run the system behind a reverse proxy that rewrites the paths.
The change removes the path prefix, and updates the docker templates to reflect the change. This will require a migration for existing systems.
This is necessary as we use zookeeper to orchestrate first-time startup of the services, to ensure that the database is properly migrated by the control service before anything else is permitted to start.
Roll back to JDK 21 for now, and make Java version configurable in the root build.gradle
The project has run into no less than three distinct show-stopping bugs in JDK22, across multiple vendors, and gradle still doesn't fully support it, meaning you need multiple JDK versions installed.
Since the performance fix in 3359f72239 had a huge positive impact without reducing result quality, it's possible to remove the QueryBranchWalker and associated code.
Previously, in an experimental change, only the first paragraph was indexed, intended to reduce the amount of noisy tangential hits. This was not a good idea, so the change is reverted.
This functionality fell into disrepair some while ago. It's supposed to allow non-mandatory search terms that boost the ranking if they are present in the document.
The change set cleans up the data model for the term-level data. This used to contain a bunch of fields with document-level metadata. This data-duplication means a larger memory footprint and worse memory locality.
The ranking code is also modified to not accept SearchResultKeywordScores, but rather CompiledQueryLong and CqDataInts containing only the term metadata and the frequency information needed for ranking. This is again an effort to improve memory locality.
The sign of the counter is used to indicate whether a term has appeared as title. Until it's seen in the title, it's provisionally saved as a negative count.
We no longer break the query into "sets" of search terms and need to adapt the code to not use this assumption.
For the API service, we'll simulate the old behavior to keep the API stable.
For the search service, we'll introduce a new way of calculating positions through tree aggregation.
The code would always re-initialize the static ngramLexicon and rdrposTagger fields with new instances even if they were already instantiated, leading to a ton of unnecessary RAM allocation.
The modified behavior checks for nullity before creating a new instance.
Seems to work, tests are green and initial testing finds no errors. Still a bit untested, committing WIP as-is because it would suck to lose weeks of work due to a drive failure or something.
This addresses the relatively common case where the graph consists of two segments, such as x y, z w; in this case we want an output like (x_y) (z w | z_w) | x y (z_w). The generated output does somewhat pessimize a few other cases, but this one is arguably more important.
Roll back to JDK 21 for now, and make Java version configurable in the root build.gradle
The project has run into no less than three distinct show-stopping bugs in JDK22, across multiple vendors, and gradle still doesn't fully support it, meaning you need multiple JDK versions installed.
Since the performance fix in 3359f72239 had a huge positive impact without reducing result quality, it's possible to remove the QueryBranchWalker and associated code.
Previously, in an experimental change, only the first paragraph was indexed, intended to reduce the amount of noisy tangential hits. This was not a good idea, so the change is reverted.
This functionality fell into disrepair some while ago. It's supposed to allow non-mandatory search terms that boost the ranking if they are present in the document.
The change set cleans up the data model for the term-level data. This used to contain a bunch of fields with document-level metadata. This data-duplication means a larger memory footprint and worse memory locality.
The ranking code is also modified to not accept SearchResultKeywordScores, but rather CompiledQueryLong and CqDataInts containing only the term metadata and the frequency information needed for ranking. This is again an effort to improve memory locality.
The sign of the counter is used to indicate whether a term has appeared as title. Until it's seen in the title, it's provisionally saved as a negative count.
We no longer break the query into "sets" of search terms and need to adapt the code to not use this assumption.
For the API service, we'll simulate the old behavior to keep the API stable.
For the search service, we'll introduce a new way of calculating positions through tree aggregation.
The code would always re-initialize the static ngramLexicon and rdrposTagger fields with new instances even if they were already instantiated, leading to a ton of unnecessary RAM allocation.
The modified behavior checks for nullity before creating a new instance.
Seems to work, tests are green and initial testing finds no errors. Still a bit untested, committing WIP as-is because it would suck to lose weeks of work due to a drive failure or something.
This addresses the relatively common case where the graph consists of two segments, such as x y, z w; in this case we want an output like (x_y) (z w | z_w) | x y (z_w). The generated output does somewhat pessimize a few other cases, but this one is arguably more important.
The previous behavior would listen to too many changes, and based on zookeeper and not curator assumptions about behavior, add an additional monitor on each invocation of each monitor, (which always trigger on service state changes), leading to each monitor re-registering and effectively doubling monitors in numbers whenever a service stopped or started, which in turn meant a lot of bizarre thrashing behavior even on changes in services that don't explicitly talk to each other.
This re-registering behavior is no longer done.
Netty and GRPC by default spawns an incredible number of threads on high-core CPUs, which amount to a fair bit of RAM usage.
Add custom executors that throttle this behavior.
Add the ability to indicate to the search service that a request is malicious, and to poison the results by providing randomly reorered old results instead.
Look, this will make the git history look funny, but trimming unnecessary depth from the source tree is a very necessary sanity-preserving measure when dealing with a super-modularized codebase like this one.
While it makes the project configuration a bit less conventional, it will save you several clicks every time you jump between modules. Which you'll do a lot, because it's *modul*ar. The src/main/java convention makes a lot of sense for a non-modular project though. This ain't that.
Cleaning out a lot of old junk from the code, and one thing lead to another...
* Build is improved, now constructing docker images with 'jib'. Clean build went from 3 minutes to 50 seconds.
* The ProcessService's spawning is smarter. Will now just spawn a java process instead of relying on the application plugin's generated outputs.
* Project is migrated to GraalVM
* gRPC clients are re-written with a neat fluent/functional style. e.g.
```channelPool.call(grpcStub::method)
.async(executor) // <-- optional
.run(argument);
```
This change is primarily to allow handling ManagedChannel errors, but it turned out to be a pretty clean API overall.
* For now the project is all in on zookeeper
* Service discovery is now based on APIs and not services. Theoretically means we could ship the same code either a monolith or a service mesh.
* To this end, began modularizing a few of the APIs so that they aren't strongly "living" in a service. WIP!
Missing is documentation and testing, and some more breaking apart of code.
The change adds a hostname validation step to remove endpoints from the ZkServiceRegistry when they do not resolve. This is a scenario that primarily happens when running in docker, and the entire system is started and stopped.
The warmup would sometimes crash during a cold start-up, because it could not get an API. Changed the warmup to just create a GrpcSingleNodeChannelPool for the node.
The query service delegates and aggregates IndexDomainLinksApiGrpc
messages to the index services. The query client was accidentally
also doing this, instead of talking to the query client.
Fixed so it correctly talks to the query client and nothing else.
Adds new ways to configure the bind and external IP addresses for a service. Notably, if the environment variable WMSA_IN_DOCKER is present, the system will grab the HOSTNAME variable and announce that as the external address in the service registry.
The default bind address is also changed to be 0.0.0.0 only if WMSA_IN_DOCKER is present, otherwise 127.0.0.1; as this is a more secure default.
The previous code made an incorrect assumption that all routes refer to the same node, and would overwrite the route list on each update. This lead to storms of closing and opening channels whenever an update was received.
The new code is correctly aware that we may talk to multiple nodes.
To avoid having to either hard-code or manually configure service addresses (possibly several dozen), and to reduce the project's dependency on docker to deal with routing and discovery, the option to use [Zookeeper](https://zookeeper.apache.org/) to manage services and discovery has been added.
A service registry interface was added, with a Zookeeper implementation and a basic implementation that only works on docker and hard-codes everything.
The last remaining REST service, the assistant-service, has been migrated to gRPC.
This also proved a good time to clear out primordial technical debt from the root of the codebase. The 'service-client' library has been taken behind the barn and given a last farewell. It's replaced by a small library for managing gRPC channels.
Since it's no longer used by anything, RxJava has been removed as a dependency from the project.
Although the current state seems reasonably stable, this is a work-in-progress commit.
In the scenario where an operator
* Performs a new crawl from spec
* Doesn't load the data into the index
* Recrawls the data
The recrawl will not find the domains in the database, and the crawl log will be overwritten with an empty file,
irrecoverably losing the crawl log making it impossible to load!
To mitigate the impact similar problems, the change saves a backup of the old crawl log, as well as complains about this happening.
More specifically to this exact scenario however, the parquet-loaded domains are also preemptively inserted into the domain database at the start of the crawl. This should help the DbCrawlSpecProvider to find them regardless of loaded state.
This may seem a bit redundant, but losing crawl data is arguably the worst type of disaster scenario for this software, so it's arguably merited.
To help services start faster, the blacklist will no longer block until it's loaded. If such a behavior is desirable, a method was added to explicitly wait for the data.
The domain blacklist blocked the start-up of each process that injected it, adding like 30 seconds to the start-up time in prod.
This change moves the loading to a separate thread entirely. For threads or processes that require the blacklist to be definitely loaded, a helper method was added that blocks until that time.
This filter currently does not distinguish itself very much from the unfiltered results, and lends the impression that the filters don't "do anything".
It may come back in some shape or form in the future, with some additional tweaking of the rankings...
Modified the DbCrawlSpecProvider to shuffle domains after loading to ensure a good mix for each crawl. This change prevents overload of crawling the same server in parallel from different subdomains or crawling big domains all at once.
Refactored the GRPC Stub Pool for better handling of channel SHUTDOWN state. Any disconnected channels are now re-created before returning the stub.
The class was also renamed to GrpcChannelPool, as we no longer pool the stubs.
Clean up the sideloading code a bit, making the Reddit sideloader use the more sophisticated SideloaderProcessing approach to sideloading, instead of mimicing StackexchangeSideloader's cruder approach.
The reddit sideloader now uses the SideloaderProcessing class. It also properly sets js-attributes for the sideloaded documents.
The control GUI now also filters the upload directory items based on name, and disables the items that do not have appropriate filenames.
Fix a bug where sideloading stackexchange files by explicitly selecting the 7z file would fail, since the 7z file would be passed along to the converter rather than the path to the pre-converted .db file.
The change deprecates the 'algorithm' field from the domain ranking set configuration. Instead, the algorithm will be chosen based on whether influence domains are provided, and whether similarity data is present.
The domain ranking code was admittedly a bit of a clown fiesta; at the same time buggy, fragile and inscrutable.
Migrating over to use JGraphT to store the link graph
when doing rankings, and using their PageRank implementation. Also added a modified version that does PersonalizedPageRank.
This change set updates the query APIs to enable the search service to add additional criteria, such as QueryStrategy and TemporalBias.
The QueryStrategy makes it possible to e.g. require a match is in the title of a result, and TemporalBias enables penalizing results that are not within a particular time period.
These options are added to the search interface. The old 'recent results' is modified to use TemporalBias, and a new filter 'Search In Title' is added as well.
The vintage filter is modified to add a temporal bias for the past.
(converter) Loader for reddit data
Adds experimental sideloading support for pusshift.io style reddit data. This dataset is limited to data older than 2023, due to licensing changes making large-scale data extraction difficult.
Since the median post quality on reddit is not very good, he sideloader will only load a subset of self-texts and top-level comments that have sufficiently many upvotes. Empirically this appears to mostly return good matches, even if it probably could index more.
Tests were written for this, but all require local reddit data which can't be distributed with the source code. If these can not be found, the tests will shortcircuit as OK. They're mostly there for debugging, and it's fine if they don't always run.
The change also refactors the sideloading a bit since it was a bit messy, and improves the sideload UX a tiny bit.
Improve the UX of the sideload GUI by sorting the results in a sensible fashion, first by whether it's a directory, then by its filename.
The change also changes the timestamp rendering to a more human-readable format than full ISO-8601.
The sideload forms didn't properly set the label 'for' property, meaning that while label tags existed, they weren't appropriately clickable.
Also removed unnecessary limits on the sideload target being a directory for stackexchange and warc. It's been possible to directly load a particular file for a while, but not allowed due to GUI limits.
Adds experimental sideloading support for pusshift.io style reddit data. This dataset is limited to data older than 2023, due to licensing changes making large-scale data extraction difficult.
Since the median post quality on reddit is not very good, he sideloader will only load a subset of self-texts and top-level comments that have sufficiently many upvotes. Empirically this appears to mostly return good matches, even if it probably could index more.
Tests were written for this, but all require local reddit data which can't be distributed with the source code. If these can not be found, the tests will shortcircuit as OK. They're mostly there for debugging, and it's fine if they don't always run.
The change also refactors the sideloading a bit since it was a bit messy.
Look at whether the property 'system.conserveProperty' is enabled when deciding he default pool size for the converter.
If true, a much more conservative default is used, limiting the risk of running out of memory.
Adds experimental support for clustering search results by e.g. domain. At a first stage, this is only enabled for the wiki and forum filters.
The commit also cleans up the UrlDetails class, which contained a number of vestigial entries.
The WARC specification says the records should transparently remove compression. This was not done, leading to the WARC typically being a bit of a gzip-Matryoshka.
The WARC specification says the records should transparently remove compression. This was not done, leading to the WARC typically being a bit of a gzip-Matryoshka.
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.
