Connecting ScyllaDB through the Cassandra integration

There is no ScyllaDB row in the connection dialog. A ScyllaDB GUI client connect starts here by choosing Cassandra, because the server speaks the CQL wire protocol on port 9042 and cassandra-driver 4.9.0 opens a session against it. What follows is which surfaces behave the same, which ones go quiet, and why the version in the header is not the product’s own number.

One integration, two servers: the ScyllaDB GUI client connect path

The connection form is the Cassandra form, with the same fields that surprise people on this engine. localDataCenter is required and sits in the open rather than behind the Advanced accordion, because the driver refuses to build a load-balancing policy without it. There is no connection string and no paste toggle: no URI convention carries a local data centre, so a paste field would parse into a connection that cannot open.

Two builds have been probed against a running instance and recorded as partial wire-compatible relatives: ScyllaDB 2026.2.4 and 2025.1.14, both on a single-node container. They behaved identically on every surface, which is why one registry entry describes both lines and only those two builds. Nothing managed, nothing clustered.

One local setup detail is easy to lose an hour to: a keyspace on the 2026.2 line must use NetworkTopologyStrategy, because that line refuses SimpleStrategy.

CREATE KEYSPACE probe
  WITH replication = {'class': 'NetworkTopologyStrategy', 'datacenter1': 1};

What reads back identically

Thirteen surfaces were called separately through the provider against ScyllaDB 2026.2.4 and, in the same pass, against cassandra:5.0.9 as the control. Eight passed on both, which covers the two things most people are here for. The editor runs CQL and returns a grid. The object browser reads the same four statements against system_schema — tables, views, columns and indexes — so tables, materialized views, column types, kinds, positions, clustering order, and index kind and target column all arrive. Columns are ordered partition key, then clustering columns, then everything else alphabetically, because system_schema.columns.position is -1 for every regular column and the server returns rows sorted by name. That is the order DESCRIBE TABLE prints and the order a WHERE clause needs, on both servers.

All 18 CQL types round-tripped byte-identically to the 5.0.9 baseline, compared field by field: bigint 9007199254740993 as a string rather than a float that has already lost its last digits, varint 123456789012345678901234567890, duration as its CQL literal 3h20m, blob as bytes, plus decimal, inet, date, time 12:00:00.123456789, list, set, map, uuid and timestamp. Normalization happens once at the driver boundary, so the grid and the export see the same values on either server.

Errors classify the same way even though the servers do not word them the same way. A missing table is unconfigured table no_such_table on ScyllaDB and table no_such_table does not exist on Cassandra. Both arrive as the recognised query error, because the classifier reads the driver’s numeric protocol code and never the server’s sentence.

Two things are absent for the integration rather than for the server, and they are absent on both. There is no EXPLAIN — the keyword is not in CQL’s grammar, so the button and the tab are not rendered rather than rendered dead. And Agent AUTO mode ends engine-unsupported, because a tool-using run needs a database-native read-only statement path, and that exists on PostgreSQL, SQLite and DuckDB only. Agent PLAN mode opens on the connection as it does everywhere: toolless, executing nothing, drafting a statement for a person to run.

The virtual keyspace that is not there

The entire delta between the two servers has one cause. ScyllaDB publishes no system_views keyspace. system.local answers, system_schema.* answers, system.size_estimates answers. system_views.clients, .queries, .caches, .system_logs and .disk_usage do not exist to be denied.

The first implementation found that out by reading the refused keyspace name out of the server’s error message, and that was the wrong instrument. Four different mistakes come back as the same protocol code 8704 with the driver’s own keyspace and table properties undefined: an absent keyspace, and a typo in a keyspace, table or column name. A build that rephrased any of those sentences would have stopped matching and taken five panels with it.

The discriminator is a catalog read now. On each successful connect the provider reads system_virtual_schema.keyspaces once — measured at 4.5 ms against Cassandra 5.0.9 and 1.6 ms against ScyllaDB — and resolves not a boolean but a reason: the catalog lists no system_views, there is no system_virtual_schema catalog at all, or the connected role may not read it. Those are three different problems that send a reader to three different places, so they are three different sentences.

It has to be that catalog and not system_schema. A virtual keyspace does not appear in system_schema.keyspaces on either server, so keying on that would have blanked all five panels on the engine that answers them. With the reason known, a monitoring refresh on ScyllaDB sends three fewer statements than the catch-the-refusal version did, because the reads that need system_views are never sent.

Which panels report absence, and with what reason

ScyllaDB publishes no system views keyspace, so the performance panel and the active-session panel report absence with their reason rather than rows, the connection count is omitted from the overview entirely, and the version on screen is the compatibility number the server publishes rather than its own release. If you need cache-hit ratios and a running-statement list from the monitoring dashboard, this connection will not give them to you, and it will say so in words rather than draw a zero.

An empty array is a claim: it says the engine looked and measured nothing. So a panel whose only source is a virtual table is left out of the payload with its reason recorded beside it, and the tab renders that sentence where the figures would have been. Before that change the Sessions tab on ScyllaDB read Active 0 / Idle 0 / Wait 0 / No active sessions found for a question the build cannot answer.

Panel	Cassandra 5.0.9	ScyllaDB 2026.2.4
Overview	version, uptime, table and index counts, connections	same, minus the connection count
Performance	cacheHitRatio from system_views.caches	refused, naming the absent catalog
Active sessions	the running statements	refused, naming system_views.queries
Table, index and storage statistics	refused, on both, with their own measurements	refused, identically
Slow queries	empty, with the reason in the panel label	empty, same reason

The bottom two rows are the point of the table. Those four panels behave the same on both servers, because their cause is Cassandra’s own: there is no honest row count and no honest size to report, and no aggregate of finished statements readable from CQL at all. The compatibility story is narrower than it looks from a distance.

The omitted connection count is worth one more sentence. activeConnections is an optional field and the provider leaves the key out rather than sending a 0, so the overview renders not published rather than 0 connections, and the agent’s health reading receives null rather than a number about a server nobody measured.

One difference shows in the tree rather than the dashboard: ScyllaDB backs a secondary index with a materialized view, so an index appears in system_schema.views, which the tree reads, and not in system_schema.tables, which the overview’s table count reads. On a keyspace of three tables and one index, the tree lists four objects and the count says three.

Why the version shown is not the product version

The header reads Apache Cassandra 3.0.8 on a ScyllaDB 2026.2.4 server, and that is not a bug in the panel. system.local.release_version is the compatibility number the server publishes, and it is the field this provider reads. The product’s own version lives in system.versions, which this provider does not read.

That is a confusing figure for a human reading a version string, so it is published as what it is rather than swapped for a friendlier one, and the caveat travels with it on the engine pages. The capability declarations behind those panels are data rather than layout, which is why a panel here can report absence with a sentence rather than fail the screen it sits on.