Wide-column databases punish people who learn them the way they learned relational databases. You cannot design a schema and then work out the queries, because the partition key determines what is possible and everything else follows from it. Query patterns come first, the schema is derived from them, and denormalisation is the expected answer rather than a failure of design. Teams that do not internalise that build a Cassandra or ScyllaDB cluster that works in development and collapses in production the first time somebody needs a query the partition key does not support.
ScyllaDB University is the best free explanation of this I have found. The catalogue is free, ungated and structured properly. ScyllaDB Essentials covers architecture and the core model. The data modelling and application development course, S201, is the centrepiece and the reason to be here.
Operations, S301, covers running a cluster in anger. There are courses on drivers, on the DynamoDB-compatible Alternator interface, and the long-running Mutant Monitoring System series that threads a single fictional scenario through a large body of material. That last device deserves specific praise. Teaching databases through a consistent, slightly absurd narrative, in this case a system for tracking mutants with superpowers, sounds like a gimmick and turns out to solve a real pedagogical problem.
It gives every example a shared context, so each new concept lands in a schema you already understand rather than in a fresh abstract table called users. The labs are hands-on against a real cluster, which is where the learning happens. Reading about how a partition key distributes data across nodes is one thing. Creating a badly chosen one, watching a single node take all the traffic, and then fixing it is a different and much more durable kind of knowledge.
ScyllaDB University LIVE runs twice a year with instructor-led foundations and optimisation tracks, also free, and the recordings persist in the catalogue afterward. Now the context that anybody investing time here should have. In December 2024 ScyllaDB moved from AGPL to a source-available licence. ScyllaDB OSS 6.2 was the final open source AGPL release, the separate open source and enterprise editions were unified into a single ScyllaDB Enterprise stream from the 2025.1 release, and a free tier of the full-featured product is available to the community with paid licensing required past a certain cluster size.
The company published its reasoning and it is a familiar argument, made by several infrastructure companies in recent years, about cloud providers capturing value from open source work. Whether you find that persuasive is a separate question from whether it affects you. Practically it means the thing you learn here is no longer open source software, there is a size threshold beyond which you are a paying customer, and the community edition exists at the vendor's discretion rather than under a licence guaranteeing it. That is worth knowing before you standardise on it.
The good news for learners is that the data modelling knowledge is not licence-dependent. Partition key design, clustering column ordering, the consequences of wide partitions, tombstones and how deletes actually behave, consistency level selection and its effect on latency and availability: all of that applies identically to Apache Cassandra, which remains Apache-licensed and widely deployed. You can take ScyllaDB's course and apply it to Cassandra with almost no translation. That makes this free material considerably more valuable than a vendor-specific course usually is.
The bias is present but mild. ScyllaDB's pitch is that it is a faster, more efficient Cassandra-compatible implementation, and the material returns to that point. Benchmarks produced by a vendor comparing itself to a competitor should always be read with suspicion, and the performance claims here are no exception. The underlying engineering argument, that a C++ shard-per-core architecture avoids the JVM garbage collection pauses that plague large Cassandra clusters, is real and well established.
The specific numbers are marketing. The absence of a proctored certification is a limitation on the credential side. You get completion certificates you can post to LinkedIn, which is not nothing and is not a credential. Four point zero.
Excellent, genuinely free, hands-on database training with data modelling content that is the best free explanation of the subject available and transfers directly to Cassandra, with a licence change in late 2024 that changes the calculus about building on the product rather than about learning from the course.