July 24th, 2024 by admin
(2024-08-03: This post got 10K+ views on the front page of Hacker News to see the followup discussion go here.)
It’s been 2 years since I worked full time in kdb+ but people seem to always want to talk to me about kdb+ and where I think it’s going, so to save rehashing the same debates I’m going to put it here and refer to it in future. Please leave a comment if you want and I will reply.
Let’s first look at the use cases for kdb+, consider the alternatives, then which I think will win for each use-case and why.
Use Cases
A. Historical market data storage and analysis. – e.g. MS Horizon, Citi CloudKDB, UBS Krypton (3 I worked on).
B. Local quant analysis – e.g. Liquidity analysis, PnL analysis, profitability per client.
C. Real-time Streaming Calcuation Engines – e.g. Streaming VWAP, Streaming TCA…
D. Distributed Computing – e.g. Margin calculations for stock portfolios or risk analysis. Spread data out, perform costly calcs, recombine.
Alternatives
Historical Market Data – kdb+ Alternatives
A large number of users want to query big data to get minute bars, perform asof joins or more advanced time-series analysis.
- New Database Technologies – Clickhouse, QuestDB.
- Cloud Vendors – Bigquery / redshift
- Market Data as a Service
Let me tell you three secrets, 1. Most users don’t need the “speed” of kdb+. 2. Most internal bank platforms don’t fully unleash the speed of kdb+. 3. The competitors are now fast enough. I mean clickbench are totally transparent on benchmarking..
Likely Outcome: – Kdb+ can hold their existing clients but haven’t and won’t get the 2nd tier firms as they either want cloud native or something else. The previous major customers for this had to invest heavily to build their own platform. As far as I’m hearing the kdb cloud platform still needs work.
Local Quant Analysis – Alternatives
- Python – with DuckDB
- Python – with Polars
- Python – with PyKX
- Python – with dataframe/modin/….
Now I’m exaggerating slightly but the local quant analysis game is over and everyone has realised Python has won. The only question is who will provide the speedy add-on. In one corner we have widely popular free community tools that know how to generate interest at huge scale, are fast and well funded. In the other we have a niche company that never spread outside finance, wants to charge $300K to get started and has an exotic syntax.
Likely Outcome: DuckDB or Polars. Why? It’s free. People at Uni will start with it and not change. Any sensible quant currently in a firm will want to use a free tool so that they are guaranteed to be able to use similar analytics at their next firm. WIthout that ability they can only go places that have kdb+ else face losing a large percentage of their skillset.
Real-time Streaming / Distributed Computing
These were always the less popular cases for kdb+ and never the ones that “won” the contract. The ironic thing is, combining streaming with historical data in one model is kdbs largest strength. However the few times I’ve seen it done, it’s either taken someone very experienced and skillful or it has become a mess. These messes have been so bad it’s put other parts of the firm off adopting kdb+ for other use cases.
Likely Outcome: Unsure which will win but not kdb+. Kafka has won mindshare and is deployed at scale but flink/risingwave etc. are upcoming stars.
Summary
Kdb+ is an absolutely amazing technology but it’s about the same amazing today as it was 15 years ago when I started. In that time the world has moved on. The best open source companies have stolen the best kdb+ ideas:
- Parquet/Iceberg is basically kdb+ on disk format for optimized column storage.
- Apache Arrow – in-memory format is kdb+ in memory column format.
- Even Kafka log/replay/ksql concept could be viewed as similar to a tplog viewed from a certain angle.
- QuestDB / DuckDB / Clickhouse all have asof joins
Not only have the competitors learnt and taken the best parts of kdb+ but they have standardised on them. e.g. Snowflake, Dremio, Confluent, Databricks are all going to support Apache Iceberg/parquet. QuestDB / DuckDB / Python are all going to natively support parquet. This means in comparisons it’s no longer KX against one competitor, it’s KX against many competitors at once. If your data is parquet, you can run any of them against your data.
As many at KX would agree I’ve talked to them for years on issues around this and to be fair they have changed but they are not changing quick enough.
They need to do four things:
- Get a free version out there that can be used for many things and have an easy reasonable license for customers with less money to use.
- Focus on making the core product great. – For years we had Delta this and now it’s kdb.ai. In the meantime mongodb/influxdb won huge contracts with a good database alone.
- Reduce the steep learning curve. Make kdb+ easier to learn by even changing the language and technology if need be.
- You must become more popular else it’s a slow death
This is focussing on the core tech product.
Looking more widely at their financials and other huge costs/initiatives such as AI and massive marketing spending, wider changes at the firm should also be considered.
Author: Ryan Hamilton