As in Power BI, a dashboard has a data mode. ingestia.bi has two: Import (table extract, zero cost when viewing) and DirectQuery (live query with cache). The cost ceiling per query applies to every query. Examples use the synthetic workspace Comércio Aurora Ltda.
Import (table extract)#
The extract-based Import mode is available to all workspaces. Older dashboards, with one query per visual, move to the extract in an assisted way, dashboard by dashboard (see the note below).
An older dashboard with one query per visual?
Move the visuals to the table extract, one by one, checking the numbers before and after: Data → Import model → Migrate visuals. See Migrate visuals to the Import model.
- What it is — the dashboard keeps an extract of each table it uses (a copy of the datalake table). Visuals are sliced from that extract, without querying the datalake again.
- When to use it — everyday dashboards that don't need real time; it's the most common case and the cheapest.
- How it works —
- Refresh now (or the dashboard's schedule) rebuilds the extracts: one read per table, no matter how many visuals the dashboard has. It's also what brings a new column added to the table into the dashboard.
- In the editor, swapping or adding a field on a visual shows up right away: the visual is sliced from the extract that's already stored, with no new query and no wait for the next refresh.
- A visual with its own query (hand-written SQL, from before extracts) keeps querying on refresh; migrate it to the extract when you can.
- Prerequisites — a configured datalake. Without a datalake, the platform runs in demo mode.
- How to validate — click Refresh now, change a field on a visual in the editor and see the number change without waiting; the dashboard footer shows the date of the last refresh.
- Limits and caveats — the extract is a snapshot: whatever changed in the table after it only shows up on the next refresh. Distinct counts don't add up across parts (e.g., unique users per day don't give the month's unique users) — use the right measure for the level the visual shows.
- Sample data (synthetic) — "Revenue by state" (SP=250, RJ=90, MG=140)
sliced from the
vendasextract. - Evidence — extract slicing, per-table extraction plan and before/after migration covered by a deterministic test suite.
DirectQuery (live)#
- What it is — the dashboard queries the source live on every view, with a cache to reduce cost.
- When to use it — when the data must reflect the current state, accepting a cost per query (on a cache miss).
- How it works — the DirectQuery engine queries live with a 10-min managed cache; billing happens on a cache miss. A widget failure becomes an error card and never takes the dashboard down.
- Prerequisites — a configured datalake; a managed cache for caching (without it, caching is a no-op).
- How to validate — set the
rtmode, open the dashboard twice within 10 min and confirm the 2nd read comes from the cache (no new cost). - Limits and caveats — the latency of each query depends on the table's volume
and on whether the result is already cached. There's no explicit timeout or cancellation of
queries in the datalake — the protection is the function's
maxDuration+ a reaper for stuck runs. - Sample data (synthetic) — 1st read queries (cache miss, billed); 2nd read within 5 min (cache hit, zero cost).
- Evidence — live query with cache (billing only on cache miss) and per-widget failure isolation covered by a deterministic test suite.
Query cache#
- What it is — a cache of query results (10 min), of the access check (60s) and of the catalog, with O(1) invalidation by version.
- When to use it — automatic in DirectQuery and on repeated reads.
- How it works — key by workspace + version + SQL hash; invalidation by version bump; in-memory fallback when there's no managed cache.
- Prerequisites — a managed cache for the shared cache; without it, it falls back to the local cache (no-op across instances).
- How to validate — repeat the same query within 10 min and confirm the cache hit; publish a new version and confirm the invalidation (miss on the next access).
- Limits and caveats — without the managed cache, the cache isn't shared across
instances; hit rate and latency gains in production are
NOT MEASURED. - Sample data (synthetic) — hash
sql:a1b2served from the cache for 10 min. - Evidence — cache by workspace + version + SQL hash, with O(1) invalidation by version and in-memory fallback, covered by a deterministic test suite.
Incremental refresh#
- What it is — recalculating only the recent period of the snapshot instead of redoing the whole history.
- When to use it — visuals with years of history where only the recent days change.
- How it works — opt-in per visual; eligibility is restrictive and
guarantees that the incremental result is identical to the full one (same
rows, same order) — at the slightest doubt, it does a full refresh. The
recentDayswindow is calculated in the dashboard's time zone. - Prerequisites — a query that can be partitioned by period and sorted ascending by the period column.
- How to validate — turn on incremental on a time-series visual and confirm that the incremental refresh number matches the full refresh number.
- Limits and caveats — it only applies when the identity conditions are met; otherwise it does a full refresh (to keep the number safe).
- Sample data (synthetic) — fixed 2024–2026 history + recalculation of the last 7 days.
- Evidence — restrictive eligibility and incremental==full identity covered by a deterministic test suite.
Cost guard (FinOps)#
- What it is — a bytes ceiling per query and billing by bytes with a minimum per query, in reais.
- When to use it — automatic on every query and dry run.
- How it works — a ceiling on volume read (1 GiB by default) on every query;
a hard
maxRowsof 50,000 (never unlimited); a minimum cost per query/table (10 MB) with USD→BRL conversion; atomic, race-safe debit of credits. - Prerequisites — a configured datalake for real billing; a workspace label on the billing export for per-customer allocation.
- How to validate — run a query and confirm, in the audit log, the recorded
costBrlandbytesBilled; try to exceed 1 GiB and confirm it's blocked. - Limits and caveats — per-customer reconciliation depends on the
workspacelabel in the billing export (still being rolled out); the recordedcostBrlis an estimate — the amount charged is the one on the invoice. - Sample data (synthetic) — a 12 MB query → billed at the minimum;
costBrlrecorded in the audit log. - Evidence — volume-read ceiling per query, hard
maxRows, minimum cost per query and atomic credit debit covered by a deterministic test suite.
Evidence: extract, cache and cost-guard engines covered by a deterministic test suite. DirectQuery latency, throughput and billing in the real cloud have not been measured yet — so this page does not promise performance numbers. Without the datalake configured, the platform runs in declared demo mode.