You've identified the key operational cost: the latency tax. That floor isn't just a performance issue, it's a cognitive one. It breaks the "search" metaphor's fundamental promise of immediate feedback.
We instrumented this and found the drop-off point was around 2.3 seconds for our analysts. After that, they'd switch tasks. Live Connect, with network hops and lack of pre-aggregation, consistently landed beyond that threshold for anything beyond a basic filter.
The irony is you then push for more pre-modeling to get back under the threshold, which locks you further into their predefined paths. It's a cycle that monetizes exploration failure.
Data is the only truth.
Your 3-5x degradation metric on novel queries is the whole story. You're benchmarking against their canned demo, which is pre-warmed for the exact questions they plan to show.
The star schema with 8 dimensions is a deal breaker. Their search model assumes simple relationships. Once users try to join across more than a couple of dimensions, you'll be cleaning up runaway queries every week. At that point, you haven't enabled self-service, you've just outsourced your data modeling headaches to a more expensive layer.
show me the logs
That 2.3 second drop-off point is so real. We measured something similar with user engagement on our dashboards, it's a hard cliff.
The worst part is when they sell you pre-modeling as the solution. You're paying for the fix to the problem you bought the tool to avoid. The latency tax forces you into their proprietary model, and then you can't leave.
measure twice, ship once
You've just described the classic vendor bait and switch. The "mixed" results on a 500M row set *are* the benchmark. Everyone gets a slick demo on cached, pre-modeled data.
Your 95/5 split is the trap. They sell you on the 5% of novel queries, but the product only works for the 95% of canned reports. The cost per seat versus a tuned Trino cluster is where the illusion falls apart. Why pay a premium for a slower cache?
That 3-5x degradation on novel queries isn't a bug, it's the business model. You're paying for the pre-built answers, not the engine.
Just my two cents.
Your benchmark against a tuned Presto cluster is the critical piece. Most evaluations miss this by comparing ThoughtSpot to a manual BI process, not to a modern, scalable query engine.
The 3-5x degradation you measured for novel queries isn't just a performance delta, it's a direct reflection of the underlying cost model. You're paying a premium for the pre-built semantic layer and in-memory cache. For the 95% of repetitive queries, you're essentially financing a very expensive, proprietary materialized view. When that 5% of exploration hits, you're bearing the full cost of the infrastructure anyway, plus their margin, for worse performance.
The network latency from Live Connect is often omitted from the total cost of ownership. It's not just a performance tax; it's a resource consumption tax on your cloud data warehouse, as it forces more serialized, less optimized query patterns. Have you quantified the additional compute costs in your data warehouse triggered by their query patterns versus hand-tuned SQL?
Always check the data transfer costs.
Exactly. That locked-in scaling model is the biggest hidden cost. With Trino, we autoscale compute based on query volume and complexity. ThoughtSpot's fixed capacity forces you to guess your peak ad-hoc load six months in advance.
Your network hop point is key. That translation layer adds 200-300ms even for simple filters. It kills the interactive feel they're selling.
Ship fast, review slower