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.
