I wonder if there's a design decision documented somewhere that makes the existing graph databases like Neo4j, etc. not good enough for Youtrack's use case.
Neo4j is a great DB but their license price is egregious for enterprise customers. A few years ago I was involved in negotiating a contract for a small/medium size kubernetes deployment (think around 25 cores) and the annual price was more than the salary of a senior SWE full-time equivalent. See this page for an idea of their prices in 2018: https://blog.igovsol.com/2018/01/10/Neo4j-Commercial-Prices....
Also embedding Neo4j is not possible, that seems to be the killer feature for YouTrackDB, they even shade dependencies so it’s like a no deps Java library for your application.
> small/medium size kubernetes deployment (think around 25 cores
That's ~1 machine. 1 SWE for a database isn't egregious, databases provide huge value, but for that little performance, that's crazy.
I can only assume as core count has blown up over the last 10 years, the pricing has somewhat diminished, but still, I'd be expecting a heck of a lot more capacity for 1 SWE.
I can tell you from quotes that I've seen, that compared to the 2018 prices listed on the page, end-of-2022 pricing was slightly up per core.
We were already well on our way of dumping Neo4J due to performance/operational/architectural reasons at the point, but seeing the quote solidified it.
Not just that, if a database company has both a community edition and enterprise then it’s likely the enterprise will get many new features that the community edition will never get.
Didn't Neo4J pivot away from being a boring embedded DB which you point at a path and then traverse through Node objects, and decide to become some kind of paid platform with a client-server protocol and proprietary query DSL?
I remembered it from a uni course (early 10s?) a few years ago for a use-case we didn't end up pursuing, but I wasn't hugely comfortable with investing effort into what I saw.
I've always been curious about graph DBs and dabbled a bit in them, but for those who have more extensive experience in them-- are they really worth it? Is it that for small scale SQL is better and graph DBs really only matter at scale, or for specific use cases with highly connected data?
I worked in a startup whose value proposition was large derived from using graph data and graph databases under the hood. The main benefit is, as the repo even states "Fast data processing: Links traversal is processed with O(1) complexity."
So, technically, you can do deep traversals quicker. A few notes:
1. Few use cases truly need low-latency deep traversal on realtime data (>5 hops or deeper). There are some well known ones like fraud detection in payment processing and, possibly, social media recommendation engines. But I am not even sure how latest social media engines work, and whether they still rely on graph DBs.
2. However, in practice the advantage is often marginal. With modern analytical databases, or even an optimized PostgreSQL (ltree, materialized views, pgrouting, pg_duckdb, etc), you often get more than good enough performance. In addition those traditional SQL DBs scale with hardware more easily than graph databases. So, you always use the lever: "Throw more hardware at it."
3. Even Graph DBs don't get good traversal performance under all conditions without hand tuning. For example, there is the"super node" issue, a node with an abnormally high number of connections (edges).
4. The ecosystem of a PostgreSQL and other popular DBs is just unbeaten. With graph DBs, you often prematurely put yourself in a corner that you don't want to be in.
Hence, my recommendation. Unless you are really sure that a graph DB is the right fit for your use case, start with something else, and go the graph db way when you have established a true need.
I worked for the research branch of a children's hospital. At one point, I was brought into vendor evaluations a department wanted to buy. Their tool had something to do with visualizing protein interactions, which was highly networked. It was basically a Neo4j database with a React UI on top of it.
I'm a bit of a Forensic Files nerd, so I worked on a little side project a while back to pull the transcripts from several episodes, use entity recognition to categorize people, places, things, etc, and load them into a Neo4j database (via Cypher queries). It turns out there's something called the POLE data model[1] that can be used by law enforcement to help solve crimes. You load all the details into a graph database and evaluate the relationships to aid in solving crimes. I suppose you could argue a criminal investigation is essentially graph traversal.
I have had very bad experiences with graph dbs at scale. To the point where I will never again work on a project based on a graph db. YMMV, but I'm done with them forever.
Most graph databases only offer an advantage in query language, as it allows for more ergonomic graph traversal compared to e.g. recursive CTEs in SQL. However with SQL/PGQ, the same query ergonomics are coming to traditional databases.
If you look under the hood there is usually nothing special about graph databases that will make them more performant. If you lay out a query plan side-by side between e.g. Postgres and Neo4J, they will look identical, just that the leaf-nodes in Postgres will be a table-scan, while in Neo4J they will be either a vertex-scan or a edge-scan (which can both be seen as special cases of a table-scan).
As someone that has worked a lot with graph databases in the past, I'd largely recommend not using them. The price you pay in terms of worse ecosystem and less battle-tested maintenance tooling is not worth it to just have better syntax.
Graph query languages like Cypher are great, but I am wholly unconvinced by the concept of a dedicated, general-purpose "graph database".
IMO, you're better off just using Postgres/etc, modeling your graphs there, and pulling in subsets of your graph for in-memory analysis. This is for the 99% of enterprises that aren't doing online streaming graph analysis, and the other 1% should probably figure out something tailored* to their specific business model.
* Graph algorithms are more accessible than ever with GenAI code, and efficiently modeling a graph is trivial (it's just structs with pointers to other edges/nodes, plus its nice to have full control over the memory layout).
We had general-purpose graph databases before graph databases became a thing. I'm talking about relational databases. Values are the vertices, tuples are the edges, FK constraints are inclusion dependencies. In fact, n-ary relations means a hypergraph database, not just a binary graph db.
Many recent graph databases are built like this. If you compare ladybugdb vs duckdb, the internals are very similar.
The main innovation is the "REL table". You can think of it as a many-to-many relationship table on steroids with optimizations at the storage layer and join algorithms.
Definition of what makes something a graph database (apart from the query language) is contested. But we seem to be moving towards: build a reasonable relational database and then add a "REL table" to it with join optimizations.
You can port files from Java to Kotlin by pressing a button, but Jetbrains has generally not done bulk ports. They leave it up to individual devs to convert a file when it's being worked on, if the diff and VCS history pollution is worth it.
Object databases routinely go away and routinely come back.
Ten years ago I worked with a database called Versant OODBMS (from Actian). I was a junior sysadmin so i was essentially administering it at a very surface level but skimming the documentation (and trying some of the samples) it was very cool that you could pick essentially any random class, implement an interface (and hence a few method) and that was it, you had a database-serializable object.
The main issue was really scaling out (as in, multiple machines) but otherwise was a really great database.
That's funny, one of my buddies from college was one of the OG Versant employees (I assume there's not another OODBMS named Versant). That was in the late 80s.
I played around with something similar before called typedb. It also does object oriented graph model which was a bit weird at first after using sql for a long time, but once it clicks there's a lot of thing that you could express surprisingly nicely
When you see 25 agent definitions, how do you know they are good? How do you know it’s better than 10 or 5 or 1? Is a project with 100 agent definitions better?
I hope this doesn't complicate the process of deploying YouTrack on-prem. That's the best part about YouTrack; historically, you can deploy on prem with no backing DB. From the outside perspective, it just stores everything to flat files.
49 comments
[ 135 ms ] story [ 1150 ms ] threadThat's ~1 machine. 1 SWE for a database isn't egregious, databases provide huge value, but for that little performance, that's crazy.
I can only assume as core count has blown up over the last 10 years, the pricing has somewhat diminished, but still, I'd be expecting a heck of a lot more capacity for 1 SWE.
We were already well on our way of dumping Neo4J due to performance/operational/architectural reasons at the point, but seeing the quote solidified it.
Ongoing enshittification risk.
I remembered it from a uni course (early 10s?) a few years ago for a use-case we didn't end up pursuing, but I wasn't hugely comfortable with investing effort into what I saw.
- custom app security
- social media
I also think cypher is a brilliant way to query a graph.
So, technically, you can do deep traversals quicker. A few notes:
1. Few use cases truly need low-latency deep traversal on realtime data (>5 hops or deeper). There are some well known ones like fraud detection in payment processing and, possibly, social media recommendation engines. But I am not even sure how latest social media engines work, and whether they still rely on graph DBs.
2. However, in practice the advantage is often marginal. With modern analytical databases, or even an optimized PostgreSQL (ltree, materialized views, pgrouting, pg_duckdb, etc), you often get more than good enough performance. In addition those traditional SQL DBs scale with hardware more easily than graph databases. So, you always use the lever: "Throw more hardware at it."
3. Even Graph DBs don't get good traversal performance under all conditions without hand tuning. For example, there is the"super node" issue, a node with an abnormally high number of connections (edges).
4. The ecosystem of a PostgreSQL and other popular DBs is just unbeaten. With graph DBs, you often prematurely put yourself in a corner that you don't want to be in.
Hence, my recommendation. Unless you are really sure that a graph DB is the right fit for your use case, start with something else, and go the graph db way when you have established a true need.
[1] https://neo4j.com/blog/government/graph-technology-pole-posi...
Edit: typo
If you look under the hood there is usually nothing special about graph databases that will make them more performant. If you lay out a query plan side-by side between e.g. Postgres and Neo4J, they will look identical, just that the leaf-nodes in Postgres will be a table-scan, while in Neo4J they will be either a vertex-scan or a edge-scan (which can both be seen as special cases of a table-scan).
As someone that has worked a lot with graph databases in the past, I'd largely recommend not using them. The price you pay in terms of worse ecosystem and less battle-tested maintenance tooling is not worth it to just have better syntax.
I like how PG19 with PGQ will kill lots of companies and startups.
Response: https://blog.ladybugdb.com/post/better-graph-database-ball/
IMO, you're better off just using Postgres/etc, modeling your graphs there, and pulling in subsets of your graph for in-memory analysis. This is for the 99% of enterprises that aren't doing online streaming graph analysis, and the other 1% should probably figure out something tailored* to their specific business model.
* Graph algorithms are more accessible than ever with GenAI code, and efficiently modeling a graph is trivial (it's just structs with pointers to other edges/nodes, plus its nice to have full control over the memory layout).
The main innovation is the "REL table". You can think of it as a many-to-many relationship table on steroids with optimizations at the storage layer and join algorithms.
Definition of what makes something a graph database (apart from the query language) is contested. But we seem to be moving towards: build a reasonable relational database and then add a "REL table" to it with join optimizations.
* Columnar storage * Compressed Sparse Row on disk * Factorized joins * ASP joins, WCO Joins
Details: https://vldb.org/cidrdb/2023/kuzu-graph-database-management-...
Ten years ago I worked with a database called Versant OODBMS (from Actian). I was a junior sysadmin so i was essentially administering it at a very surface level but skimming the documentation (and trying some of the samples) it was very cool that you could pick essentially any random class, implement an interface (and hence a few method) and that was it, you had a database-serializable object.
The main issue was really scaling out (as in, multiple machines) but otherwise was a really great database.
http://www.kevra.org/TheBestOfNext/ThirdPartyProducts/ThirdP...
[0]: https://github.com/falkordb/falkordb