> You have to handwash the dish, then put it through the dishwasher, then handwash it when it comes out.
You have never actually used a dishwasher have you.
93 karma · joined December 13, 2011
You have never actually used a dishwasher have you.
Fireplace - tooling to deploy Fireproof apps and sync data across your Tailscale network. Once all the computers you care about are on your tailnet, of course you want all the browsers on the tailnet to easily sync with one another.
Go Implementation - Fireproof bills itself as a realtime database that runs anywhere, and I want to make sure that includes inside your Go applications. This will allow your Go application to become a full-fledged reader/writer of the Fireproof ledger.
I'm excited to see what other people want to build and answer any questions.
Bleve does (optionally) support persistence, so reading/writing files is one place it does directly interact with the environment. The environment must support mmap.
There are several projects which support distributed index/search workloads with Bleve. The exact approaches vary, but they all use Bleve to perform node local operations, and coordinating this is done at a higher level by the application.
I suspect I don't understand the terminology you're using in the last question, as Bleve has no training, models or nodes.
1. Make Bleve a distributed index, or make Bleve into something that is a more direct ES competitor.
We have no plans to do this because we think that is better built at a different layer. We have hooks we introduce in certain places where we need to plug-in code that would otherwise violate the boundaries. And that is an arrangement that has worked well so far. There are multiple projects built on top of bleve that allow you to index/search across nodes.
2. Make an adapter for the XYZ key/value store.
This request goes back to the original bleve index which is serialized into a key/value abstraction layer. When users run into size/speed issues with bleve, many assume that just plugging in a faster key/value store will help. (Hey we thought that too when we built it this way)
But, we've now replaced that index scheme with a new implementation called scorch. Scorch is considerably smaller and faster, and manages it's own index on disk, without using any key/value store.
As for things that we DO plan to implement:
1. Size of the index still comes up a lot. Couchbase is a very performnace sensitive user of Bleve, so I expect they'll lead the way on this front.
2. Better (pluggable) scoring. Today our search result scoring is broken for several types of queries, and the stuff that does score right is too tightly coupled to the searching logic.
3. Overhaul index mapping. Today bleve uses a mapping object to describe how source objects/documents are indexed. One of the best ways we can simplify the mapping is to make things more explicit. I think we tried to embrace the concept of reasonable defaults, but we ended up with inheritance hierarchies that are difficult to reason about.
There are lots of miscellaneous things like adding a data type that supports IPv6, or more advanced queries (lots of variations on span queries).
However, many of those same users complained about having to operate another cluster, especially ones that weren't already using the JVM (since it was a skill set they didn't have).
So, the appeal was to offer a service that runs as a part of the Couchbase cluster. It wouldn't have to match every feature of ES, just shoot for 80/20 and customers would likely find it beneficial.
It was fortunate that Go was still growing in popularity within Couchbase at that time, and we were able to position Bleve as a true open-source component, on top of which some money-making value add could be layered.
One of the big things we're working on at the moment is improving the release process. In addition to semantic versioning of the APIs we have to think through how it applies to the binary artifacts created. We want Go modules to be supported and be a part of the solution, but we are also mindful not to break things for existing users.
Concurrent indexing is also possible, so long as you can arrange to not put duplicate document ids into batches executing concurrently.
As for usage, it is just a library, so it is typically embedded in a single process (though this can serve multiple clients concurrently).
Distributing the index across multiple nodes is done at the application level. At Couchbase we do this with bleve in a separate project called 'cbft'.
https://github.com/blevesearch/blevesearch.github.io-hugo/is...
With scorch the index size is 22MB. Query performance is comparable (and we haven't even gotten to really tuning this yet).
Merging is required and is indeed somewhat resource intensive. Bleve's current indexing approach has no segments, instead all index data is serialized into a key/value store. This approach allowed us to experiment and plug-in a variety of implementations. Unfortunately, the key/value abstraction limits the way you interact with data, so there are a number of drawbacks. One key gain we get from the segmented approach vs the key/value store approach is that we no longer need to maintain a backindex to handle updates/deletes.
Just recently we merged support a new experimental index scheme called 'scorch'. This new index scheme is designed from the ground up to reduce index size and improve performance. It features:
- a segment based approach, much like Lucene
- vellum FTS for the term dictionary - https://github.com/couchbase/vellum
- roaring bitmaps for the postings lists - https://github.com/RoaringBitmap/roaring
- and compressed chunked integer storage for all the posting details
It's still experimental at this point, but shows considerable indexing speedup, index size reduction, and similar query performance to the old index format used today.
The code for this new index scheme can be found here: https://github.com/blevesearch/bleve/tree/master/index/scorc...
The values after the / are the output values associated with the transition.
Weighted finite state transducers are not supported.
Author here, yes the observation that it is immutable is correct. One approach you can use is to have some lighter-weight representation for the most recent data. Then have one or more FSTs representing the older data. As time permits keep merging the new data and the older ones down into new larger FSTs. Merging them is straightforward since you can iterate the contents in order, which is the order you need to build new ones. In this way, it is very similar to having a WAL up front, and one or more segments backing an LSM storage.