Preview and analysis

Tabularia shows the data at every step: click a node of a flow and its result appears under the canvas, as a table or as a chart. The Viewer does the same on any datasource without building a flow: filters, computed fields, a pivot and charts, computed by the engine on the whole table, and saved as a named view that anyone with access to the folder can reopen. Nothing in either place changes the data.

At a glance

Previewing each step of a flow

Clicking a node runs the chain that leads to it and shows its first 100 rows, with the column types, in the Table tab. The count above the table says how many rows came back and whether there are more («100 rows (truncated)»); the node keeps the same count («100+ rows · 12 columns») so you can read the shape of a flow at a glance. Values are shown raw, nulls as ∅. A node that is not configured yet asks to be set, instead of showing an error.

Each click shows that step's first 100 rows: the order lines, the computed revenue, the join with the orders, the margin by category and month.

The rows are computed on the whole input: a Group by groups every row, a Join joins every row, and you see the first 100 of the result. The one exception is a source with a Development sample (first N rows or a random percentage), which previews and editor runs read instead of the full table. An administrator can give every source an automatic sample with PREVIEW__DEFAULT_SAMPLE_ROWS (0 by default: no sample); the source node then shows «automatic (first N rows)» and anyone can switch it to «off (all rows)». Scheduled runs and Run in production always read everything.

The Chart tab draws the selected node. The chart is an aggregation the engine computes on all the node's rows, not on the 100 shown: choose the kind (Bars, Lines, Area, Pie, Treemap, Scatter), the column for the X axis, an optional column that splits it into series, and a function: count, sum, average, minimum, maximum, median or distinct count. Bars keep the top N categories (20 by default), a treemap the top N too; a pie shows the five largest and, for a count or a sum, gathers every other category of the data under «Other». A pie, a treemap and a funnel draw only what is above zero: they say how many categories they leave out, and with nothing above zero they say so instead of drawing. At most five series are drawn, and the chart says how many it left out. A scatter plots one point per group, with an aggregation on each axis.

More chart kinds

After the first six, the row of chart kinds has two more groups. The first composes: Bars and line (a second measure drawn as a line over the bars), Waterfall (how each category adds to or takes from the total, which comes last), Funnel (stages from the largest down) and Heatmap (X by a second column, by, the colour by the value). The second shows one number: KPI (the aggregate of the whole table in large type; with trend, a column such as a month, its sparkline and the change of the last period against the one before), Gauge (the number on an arc, with min, max and a target) and Bullet (a bar per category against a second measure, its target). They are computed on all rows like the others, look as they do in the pixel-perfect reports, and the assistant draws them in its answers too.

The Viewer with Gauge: the average margin percentage of the sample data against a target of 30; the row of kinds above, the first six, those that compose, those of one number.

Download this node's data, in the node's panel, writes the node's whole result as CSV or Excel: a worker writes the file and the browser saves it.

The step cache

Every step's result is stored under a key made from its source and its operations. Previewing node 9 after node 8 reuses node 8's result instead of starting from the source again, and changing the last step of a long flow recomputes one step, not all of them.

• The copy never delays the answer. The preview answers first; the step's result is written afterwards by a separate, low-priority task.
• The editor shows what is cached. A node whose result is in the cache has a ring and a small icon, «Result in cache: the nodes after this one start from here».
• Large steps are not copied. A step over CACHE__MAX_STEP_ROWS rows (1 million by default) is recomputed at every preview: a copy as large as the source saves nothing. The node shows a refresh icon and the editor warns once, suggesting a development sample or an earlier filter.
• Entries expire. A step unused for CACHE__TTL_SECONDS (one day by default) is deleted. On BigQuery the cache is a set of expiring tables in a dataset of its own (BIGQUERY__CACHE_DATASET).

The Viewer does not use the step cache: its filters change at every click, and a cache of them would only fill the storage.

Cancelled previews

Moving through the nodes of a flow starts a preview at every click. Only the last one matters, so each panel (the table, the chart, the list of a column's values) keeps one preview at a time: a new one cancels the previous, whether it is still queued or already running. On the external ClickHouse engine the query is stopped on the server too, and on BigQuery the job is cancelled. A preview that takes longer than PREVIEW_TIMEOUT_SECONDS (120 by default) is stopped and reported.

The Viewer

The Viewer is a read-only analysis page over one datasource. Choose the datasource and the engine at the top (your Preferred engine, from Settings, is the default), then build the view in the side panel and press Apply:

• Filters: =, ≠, >, ≥, <, ≤, between, contains, starts with, in (a,b,…), not in, is null, is not null. Dates and times get a calendar.
• Computed fields: a name and a SQL expression in the engine's dialect, for example ricavo / chili. They can be filtered, pivoted and charted like any column.
• Pivot: Rows (dimensions), Columns (dimensions), Values and a Function (sum, average, min, max, count). Don't repeat row dimensions prints a dimension's value once per group, as in a report.
• Table and Chart tabs: the table shows the first 500 rows of the result; the chart works as in the editor, on the filtered and computed data.
• Export CSV and Export Excel download the whole current view (filters, computed fields and pivot applied). The chosen engine computes it; the file is written by a worker with Polars.

The saved view «Margin by category — month by month»: Pivot with Rows categoria, Columns anno_mese, Values margine, Function sum.

Saved views

Save view stores the current configuration, with a name, in a folder: the engine, the filters, the computed fields, the pivot, the open tab and the chart's settings. A view holds no rows: opening it tomorrow reads tomorrow's data. Saved views appear in Explore next to flows and datasources, and in the View picker of the Viewer. Saving again under the same name in the same folder updates the view; the name cannot be reused there for a view of another datasource.

Example: the margin of the sample data

The sample flow «Margin by Category and Month» publishes a datasource of the same name in Sample / Private tables, with one row per product category and month: categoria, anno_mese (202510 for October 2025), ricavo, costo, margine, chili, ordini, margine_pct.

1. Open Analysis → Viewer and choose Margin by Category and Month.
2. Filters: margine_pct < 30, then Apply. The table lists the months in which a category earned less than 30% of its revenue.
3. Computed fields: ricavo_al_kg = round(ricavo / chili, 2), revenue per kilo.
4. Remove the filter and turn the Pivot on: Rows categoria, Columnsanno_mese, Values margine, Function sum. One row per category, one column per month.
5. Switch to Chart, choose Lines, X anno_mese, series categoria, function sum of margine: the seasons of the sample data (Christmas, summer) show up in each category's line.
6. Save view, name Margin by category — month by month, in a folder of yours. Whoever can see that folder opens the same matrix, on the current data.

The Sample folder already holds four views built this way, among them «Margin watch — categories under 30%» (step 2) and «Margin by category — month by month» (step 4).

The same view: the rows without the pivot, the matrix with it, then Chart with the sum of margine by category, computed on every row.

Permissions

A view grants access to nothing: whoever opens it without access to the datasource gets a message, not the data. Every download and export is in the audit log (export.download), with the format, the file name and the sources read.

Limits and good to know

• The editor shows 100 rows per node and the Viewer 500; charts and exports read every row.
• A CSV export holds at most EXPORT__MAX_ROWS rows (5 million by default) and an Excel file the format's limit of about one million. Larger results belong in a datasource, through a flow's Output node.
• One person has at most EXPORT__MAX_RUNNING_PER_PERSON exports and documents (CSV, Excel, reports, dashboards) being prepared at once, 3 by default: past it Tabularia asks to wait for one to finish. Code — the REST API with a token, a notebook's tabularia.report() — waits for its turn instead (API).
• A pivot makes at most 500 new columns.
• A preview's numbers come from the engine you previewed with. When a flow's production engine is another one, see Engine differences.
• Joining two sampled sources loses most matches: sample only the big table.

Related

• Flows: the editor and its nodes
• Engines: which engine computes a preview
• Ask your data: questions in plain language over the same datasources
• SQL scripts: free SQL over the catalogue, with CSV and Excel export
• Configuration: the cache settings