github.com/weaviate/weaviate@v1.24.6/adapters/repos/db/lsmkv/doc.go (about) 1 // _ _ 2 // __ _____ __ ___ ___ __ _| |_ ___ 3 // \ \ /\ / / _ \/ _` \ \ / / |/ _` | __/ _ \ 4 // \ V V / __/ (_| |\ V /| | (_| | || __/ 5 // \_/\_/ \___|\__,_| \_/ |_|\__,_|\__\___| 6 // 7 // Copyright © 2016 - 2024 Weaviate B.V. All rights reserved. 8 // 9 // CONTACT: hello@weaviate.io 10 // 11 12 /* 13 # LSMKV (= Log-structured Merge-Tree Key-Value Store) 14 15 This package contains Weaviate's custom LSM store. While modeled after the 16 usecases that are required for Weaviate to be fast, reliable, and scalable, it 17 is technically completely independent. You could build your own database on top 18 of this key-value store. 19 20 Covering the architecture of [LSM Stores] in general goes beyond the scope of 21 this documentation. Therefore things that are specific to this implementation 22 are highlighted. 23 24 # Strategies 25 26 To understand the different type of buckets in this store, you need to 27 familiarize yourself with the following strategies. A strategy defines a 28 different usecase for a [Bucket]. 29 30 - "Replace" 31 32 Replace resembles the classical key-value store. Each key has exactly one 33 value. A subsequent PUT on an an existing key, replaces the value (hence 34 the name "replace"). Once replaced a former value can no longer be 35 retrieved, and will eventually be removed in compactions. 36 37 - "Set" (aka "SetCollection") 38 39 A set behaves like an unordered collection of independent values. In other 40 words a single key has multiple values. For example, for key "foo", you 41 could have values "bar1", "bar2", "bazzinga". A bucket of this type is 42 optimized for cheap writes to add new set additions. For example adding 43 another set element has a fixed cost independent of the number of the 44 existing set length. This makes it very well suited for building an 45 inverted index. 46 47 Retrieving a Set has a slight cost to it if a set is spread across multiple 48 segments. This cost will eventually reduce as more and more compactions 49 happen. In the ideal case (fully compacted DB), retrieving a Set requires 50 just a single disk read. 51 52 - "Map" (aka "MapCollection") 53 54 Maps are similar to Sets in the sense that for a single key there are 55 multiple values. However, each value is in itself a key-value pair. This 56 makes this type very similar to a dict or hashmap type. For example for 57 key "foo", you could have value pairs: "bar":17, "baz":19. 58 59 This makes a map a great use case for an inverted index that needs to store 60 additional info beyond just the docid-pointer, such as in the case of a 61 BM25 index where the term frequency needs to be stored. 62 63 The same performance-considerations as for sets apply. 64 65 # Navigate around these docs 66 67 Good entrypoints to learn more about how this package works include [Store] 68 with [New] and [Store.CreateOrLoadBucket], as well as [Bucket] with 69 [Bucket.Get], [Bucket.GetBySecondary], [Bucket.Put], etc. 70 71 Each strategy also supports cursor types: [CursorReplace] can be created using [Bucket.Cursor], [CursorSet] can be created with [Bucket.SetCursor] , and [CursorMap] can be created with [Bucket.MapCursor]. 72 73 [LSM Stores]: https://en.wikipedia.org/wiki/Log-structured_merge-tree 74 */ 75 package lsmkv