ThoughtSpot’s Data Cache is fast—but cached and live answers are not always equivalent
The 26.11 feature moves eligible Spotter and analytics queries into ThoughtSpot-managed storage. Teams should treat it as a second execution surface, not a transparent accelerator.
ThoughtSpot Cloud 26.11 makes model-level Data Cache generally available: it copies selected model tables from Snowflake or Databricks into ThoughtSpot-managed storage and automatically routes eligible queries to that cache. The attraction is straightforward—ThoughtSpot says cache hits typically return with sub-second latency and reduce warehouse compute—but the operational contract is more consequential than a speed toggle. ThoughtSpot’s Data Cache documentation shows that teams are introducing a second execution surface with its own refresh schedule, storage boundary and query semantics.
The routing is automatic; the equivalence is conditional
Users do not choose whether a query runs against the cache or the live warehouse. ThoughtSpot checks eligibility at query time and falls back to the warehouse when a date filter falls outside the cached window, a construct is unsupported or the cache is unavailable. During initial configuration, first load or a failed refresh, queries also run live; during a refresh, the most recent successful cache remains active. The documented state model therefore makes answer provenance dependent on cache state and query shape, even though the user receives one completed result either way.
The bigger rollout caveat is that the two paths can behave differently. ThoughtSpot documents case-sensitive cached handling for text that Snowflake may collate case-insensitively, 15-significant-digit handling for decimals, compact-JSON storage for VARIANT and ARRAY columns, a different result for logarithms of zero, errors on infinity values and a null yearly bucket for dates in year 0001. It also says edit_distance() is not yet supported on cached models. ThoughtSpot’s compatibility table recommends source normalization or keeping affected models live.
Start with a narrow model, not the whole estate
The first release supports SaaS clusters, Snowflake and Databricks connections, and service-account connections rather than OAuth or parameterized connections. Models with custom calendars cannot be cached, and row-level security rules that reference other tables always send queries to the warehouse. Cluster support must enable the feature, then an editor must opt in each model; ThoughtSpot explicitly warns that data is copied out of the customer warehouse into its managed storage. The prerequisites and security sections also say row- and column-level controls remain applied and each organization’s cache is isolated.
A sensible pilot is one high-volume model whose dimensions can be fully copied and whose fact table fits a bounded date window. ThoughtSpot offers 1-, 3-, 6- and 13-month windows plus custom durations, with hourly through monthly refresh schedules. Teams should record which questions hit the cache, test boundary values from the compatibility table, and compare cached versus forced-live answers before claiming equivalence. That last control is especially important for Spotter: faster natural-language answers are useful only if users are not silently crossing between execution semantics.
Finally, plan for lifecycle work. Renaming columns or adding parameters requires disabling caching first; turning it off clears cached data and run history. Cache data is not included in cluster snapshots, so it must be rebuilt after a restore. ThoughtSpot’s operating instructions make Data Cache a managed data product, not a set-and-forget optimization.
sources
- ThoughtSpot Data Cache documentationdocs.thoughtspot.com
- ThoughtSpot Cloud 26.11 release notesdocs.thoughtspot.com
comments · 0
- Databricks made ai_decide more accurate—but the model behind the SQL function can change againNov 17, 2026
- A new French open-data test finds architecture matters less once the local model reaches 3B parametersNov 17, 2026
- Genie One scheduled tasks can now email colleagues—treat the recipient list as a data-delivery surfaceNov 17, 2026