There is one caveat I would add. While you can always build just about any kind of database or data model on top of a logical KV store, the scalability and performance is almost always going to be qualitatively worse for queries than using a storage engine purpose-built for the intended type of database or data model. It is simpler to implement, which is why people do it, but you give up a lot of selectivity and locality on the query side.
The target design case for KV stores is when all your queries are essentially trivial. Layering SQL on top of it implies the expected use case is complex queries of the type where a KV store won't perform well. Good for a compatibility interface, but not performant.
I've always considered KV to be an API. Under the hood you're still using data structures managing fixed-size blocks and byte-offsets, such as LSM trees, and using binary searches to minimize disk reads (e.g. index files of very 128th 64k key block).
If you're building something more robust on top of a KV store, I'd be under the impression you're not doing so using the logical API but leveraging the underlying data structures with algorithms better suited to the task. An LSM tree is pretty decent at random seek/scan, and implementing something like JOINs is less about KV constraints as it is LSMT constraints.
For example, FoundationDB doesn't use SQLite directly but rather its B-tree implementation.
I intentionally added the word "logical" because I think it is important. All database engines are KV at the physical page/block I/O level but they almost never expose that to the user, and there is no requirement to limit the expression to what the lower level can express.
The primary limitation of KV as an API is that it can't directly express complex but elementary relationships between keys e.g. intersections when your keys are constraint types. Or graph traversal relationships, if efficiency on real hardware matters at all. Most complex databases, like SQL, are implicitly searching across relationships like this constantly. Not being able to express these types of relationships as first-class citizens is computationally expensive.
Putting a KV store under every database is a recent fad. Databases classically had more expressive but still inadequate storage architectures (e.g. first-class operations on constraint types were terrible there too), so this is a step backward in most ways -- but I understand why expediency incentivizes that approach. Kind of like discarding joins to achieve scale-out -- joins are critical operations for many things but scaling out joins requires very challenging computer science. On the flip side, a lot of the cutting edge research has been focused on jamming as much expressiveness in a singular index-organized store as possible, in as little space as possible. This has proven to be a very effective, albeit more difficult to understand.
LSM trees are a good example of an incremental hack to address limitations in more traditional approaches while not solving the underlying fundamental issues with the way the indexing structures are constructed. LSM tree write throughput is okay but not great, empirically. Similarly, the query performance is okay but not great, for reasons that are well-understood from other contexts in database engine design. You can use it to solve a specific problem but it is unclear why it should be the solution to a general problem.
KV stores, LSM trees, etc are stuck in a suboptimal local minima.
Could you elaborate on this argument? For something like a graph, I din't quite follow why a KV API is problematic or how you'd do it better. If you want to map node locality to KV locality you can certainly do that.
The core design problem for graph databases is data locality during join recursion. If you shard your data model by key, the cross-shard join operation asymptotically converges on a Cartesian product for each join iteration i.e. pathological data locality. This is why most graph databases have such poor scalability in practice.
A less intuitive approach is to shard the data model based on edge similarity measures, such that a specific key does not map to a single shard. While this obviously has poorer locality for simple key lookup, cross-shard join operations -- the most expensive operation and what actually matters -- only involve a tightly bounded number of shards and therefore have much better locality. While this was originally developed for to make the problem scale on supercomputers (at IBM Research AFAIK), it is entirely amenable to use in graph databases.
At the storage engine level, first-class support for indexing by similarity measures has different design requirements and tradeoffs than simple key lookup or ordered-tree indexing. While you could make it work on a KV store, the impedance mismatch would incur a significant performance drag. This cuts both ways; storage engines optimized for use with similarity measure designs are going to offer poor performance if you try to put b-trees or LSM-trees on top of them.
I think of FoundationDB as a low level persistence layer that you can use to build a fully functional database on top with more complex querying capability. It provides the core set of semantics such as data replication, globally ordered transactions and horizontal scalability - that are all complex to build in their own right.
If you look at CRDB or TiKV design, under the hood they are also KV-stores with SQL layered on top.
Very curious, what’s an alternative implementation strategy or more appropriate mental model?
My mental model (I am not a database engineer) is that every SQL database is fundamentally backed by a key value store. For Postgres, each value is a row of a relation, the key is the relation’s key, and the physical storage is the table file. Postgres builds b-trees on top to make it fast, but the b-tree itself is a relation, a key value store that Postgres works hard to keep sorted.
Many data models don't have obvious singular keys. The critical search relationships cannot be reduced to an order relationship on a single column a priori. Much of the most interesting research in storage models is around increasingly efficient ways of indexing complex information theoretic features across multiple columns in a single structure, with an underlying storage implementation to match.
At some level, every database contains a key-value store. For performance reasons, the hardware always has to be treated this way. Databases work at the level of blocks/pages, but those abstractions are usually hidden from users with a lot of clever logic in the middle that is more opinionated to enable optimization. That doesn't change.
An interesting and important property of search data structures is that, at the limit, a single index that can optimally satisfy all possible queries is equivalent to general AI. It is also completely intractable. Fortunately real-world queries tend to be much more limited in nature. A corollary is that the distinction between indexing, storage, and scheduling in databases is a fiction -- useful for making some things simple but not necessary in any database. In essence, the practice of treating indexing, storage, and scheduling as discrete functions in a database is the opposite extreme. There are a vast number of possible implementations between these two extremes with better properties than either in practical real-world databases.
As a general design principle for scalability and performance reasons, you want to organize your data model around a single indexing mechanic. Consequently, it is critical to maximize the expressiveness and efficiency of any particular indexing mechanic. At the limit, with a good algorithm, it is equivalent to having an index for each column, with the ability to efficiently search more features of the data and without the overhead of actually having an index for each column.
I think we are entering a new golden age of database technology where the boundaries between elements we treated as discrete are much fuzzier.
i’m also not a db engineer, but i think this is true-ish. however building and maintaining those index tables is hard and probably prone to issues if you can’t update multiple as part of the same transaction.
the other major thing you’d miss is smarter joins. the distributed databases do a lot of work to be able to push down predicates as far down as possible to the physical nodes storing the data.
though as with things like CAP theorem, there are always clever ways to to place a given implementation at various points along the spectrum, rather than just at one end or the other.
The target design case for KV stores is when all your queries are essentially trivial. Layering SQL on top of it implies the expected use case is complex queries of the type where a KV store won't perform well. Good for a compatibility interface, but not performant.