This post outlines the architecture and design decisions behind Vokter, an open-source Java library that I’ve been developing since October 2014. This does not address other projects under the ‘Vokter’ umbrella, such as HTTP servers that use the core library to manage a persistent state of Vokter on the web instead of locally, but I do hope to address them later in a future post.
What started as a simple project for a University class in Distributed Architecture, which I named Argus (from Argus Panoptes, the all-seeing giant of Greek mythology), it is currently one of my most ambitious and complex libraries, now known as Vokter (Norwegian for watcher).
Vokter is a high-performance, scalable Java library that combines Locality-Sensitive Hashing for K-Shingles, a fork of DiffMatchPatch, Bloom filters and Quartz jobs to detect changes in web documents, triggering notifications when specified keywords were either added or removed.
At a basic level, Vokter fetches web documents on a periodic basis and performs detection (comparison of occurrences between two snapshots of the same document) and matching (finding out if one of the detected differences matches a registered keyword, sending messages to attached consumers if so).
It optionally supports multi-language stopword filtering, to ignore changes in common words with no important significance, and stemming to detect changes in lexical variants. Appropriate stopword filtering and stemming algorithms are picked based on the inferred language of the document, using a N-grams Naïve Bayesian classifier.
There are two types of jobs, concurrently executed and scheduled periodically (using Quartz Scheduler): detection and matching jobs.
The detection job is responsible for fetching a new document and comparing it with the previous document, detecting textual differences between the two. To do that, the robust DiffMatchPatch algorithm is used. The matching job is responsible for querying the list of detected differences with specific requested keywords.
Harmonization of keywords-to-differences is performed by using a Bloom filter, removing differences that have a very low chance of containing the specified keywords, before using a exact comparator (character-by-character) on the remaining differences, to ensure that the detected changes contains any of the keywords.
Since the logic of difference retrieval is spread between two jobs, one that is agnostic of requests and one that is specific to the client and its keywords, Vokter reduces workload by scheduling only one detection job per watched web-page. For this, jobs are grouped into clusters, where its unique identifier is the document URL. This means that each cluster imperatively contains a single scheduled detection job and one or more matching jobs. In other words, When two clients watch the same page, only one detection job for that page is active.
Vokter was conceived to be able to scale and to be future-proof, and to this effect it was implemented to deal with a high number of jobs in terms of batching and persistence.
The clustering design mentioned above implies that, as the number of clients grows linearly, the detection logic remains independent of the clients and is only executed at a given set of triggers.
In terms of orchestration, there are two mechanisms created to reduce redundant resource-consumption, both in memory as well as in the database:
- if the difference detection job fails to fetch content from a specific URL after 10 consecutive attempts, the entire cluster for that URL is expired. When expiring a cluster, all of the associated client REST APIs receive a time-out call;
- every time a matching job is canceled by its client, Vokter checks if there are still matching-jobs in its cluster, and if not, the cluster is cleared from the workspace.
Documents, indexing results, found differences are all stored in MongoDB. To avoid multiple bulk operations on the database, every query (document, tokens, occurrences and differences) is covered by memory cache with an expiry duration between 20 seconds and 1 minute.
Persistence of detection and matching jobs is also covered, using a custom MongoDB Job Store by Michael Klishin and Alex Petrov.
The string of text that represents the document snapshot that was captured during the Reading phase is passed through a parser that tokenizes, filters stop-words and stems text. For every token found, its occurrences (positional index, starting character index and ending character index) in the document are stored. When a detected difference affected a token, the character indexes of its occurrences can be used to retrieve snippets of text. With this, Vokter can instantly show to user, along with the notifications of differences detected, the added text in the new snapshot or the removed text in the previous snapshot.
Because different documents can have different languages, which require specialized stemmers and stop-word filters to be used, the language must be obtained. Unlike the
Content-Type, which is often provided as a HTTP header when fetching the document, the
Accept-Language is not for the most part. Instead, Vokter infers the language from the document content using a language detector algorithm based on Bayesian probabilistic models and N-Grams, developed by Nakatani Shuyo, Fabian Kessler, Francois Roland and Robert Theis.
To ensure a concurrent architecture, where multiple parsing calls should be performed in parallel, Vokter will instance multiple parsers when deployed and store them in a blocking queue. The number of parsers corresponds to the number of cores available in the machine where Vokter was deployed to.
Vokter support for reading of a given
MediaType is provided by Reader modules, where raw content is converted into a clean string filtered of non- informative data (e.g. XML tags). These modules are loaded in a OSGi-based architecture, meaning that compiled Reader classes can be loaded or unloaded without requiring a reboot. When needed, usually when reading a new document or snapshot, Vokter will query for available Readers by
This same plugin-like architecture is implemented for Stemmer modules. Using a language detection prediction model, Vokter determines the most probable language of the document and queries on-demand for available Stemmers by language supported.
Despite every part of its architecture having been optimized to accommodate to a massive amount of parallel tasks, Vokter has only been used in a academic environment and has yet to be battle-tested in high-usage consumer software. If you're using Vokter in your projects, let me know! 😀
Regardless, there is always room for improvement, and I feel that there are currently two main issues that should be addressed with higher priority: 1) web crawling functionality; and 2) timeout of jobs when clients are missing.
One way to improve user experience is by integrating web crawling in Reader modules, allowing users to set their visit policy (e.g. number of nested documents accessed). Within the current architecture where there is a unique detection job per document, detection jobs must be sorted by link hierarchy.
Let's imagine an example where a job 1 watches document A and a job 2 watches document B, where A has a link to B. In this case, job 2 does not need to exist as there would be redundancy of detection jobs and index storage for document B. Instead, job 1 should trigger clients linked to A and B:
- when differences are detected in A, only clients of A are notified;
- when differences are detected in B, both clients of A & B are notified.
This implies a more optimized architecture that has the potential of significantly reducing the total number of simultaneous jobs.
Orchestration for matching jobs
After an attempt to load a new snapshot of the document fails too many times, only detection jobs are timed-out. However, the system can fail to send a response to the client as well, and there is currently no way to deprecate matching jobs when the client has "disappeared and lost interest" before canceling their jobs from Vokter. This means that a high number of active detection and matching jobs might be kept alive unnecessarily.
More on projects
Developing a knowledge-based system for Texas Hold'em outcome prediction in Erlang
Implementation of a decision-tree classifier and a disk persistence module to evaluate win-lose probabilities over a Poker Texas Hold'em game