Buyer’s guide Sourced and dated
If you want to drag fields around a Parquet or CSV file you already have on disk, with no server and no repo, use xmetric: it opens the file in a browser tab and runs the queries there. Pick Evidence or Observable Framework when the result should be a versioned report in Git, and Superset when a team needs a shared portal on a warehouse.
Our pick
xmetric
It opens the file in a browser tab with nothing to install or host, the data stays on your machine, and you pivot on shelves with Tableau-style formulas instead of rewriting a GROUP BY for every question.
Pick something else if
The output is a report that lives in Git and deploys on CI (Evidence or Observable Framework), or a team needs a shared, permissioned portal on Postgres, Snowflake or Trino (Superset or Preset).
xmetric makes this page. Every vendor fact below links to the vendor’s own page and the date we checked it.
Here is a month of events from a habit app's pipeline. Count rows per user and Android users log 19.7 habits each against 14.3 on iOS, a 38% lead. Then compare rows with distinct event ids: 307 of Android's 995 rows are repeats of an event already stored, from a client that retries uploads. Count distinct events and Android lands at 13.7, just behind iOS. Nothing about engagement changed; the table had duplicates.
Computed by the xmetric engine
The data
A habit-tracking app with 60 users on iOS, Android and web. Four weeks of events as the ingestion pipeline wrote them to Parquet, retries included.
event_id | user_id | platform | event_date | event_type |
|---|---|---|---|---|
| e000040 | u0002 | ios | 2025-09-01 | session_start |
| e000083 | u0003 | ios | 2025-09-01 | session_start |
| e000084 | u0003 | ios | 2025-09-01 | habit_logged |
| e000106 | u0004 | ios | 2025-09-01 | session_start |
Step 1
Android looks well ahead of iOS. Before you tell the mobile team, check what a row is.
Tableau
SUM(IF [event_type] = "habit_logged" THEN 1 END) / COUNTD([user_id])SQL
SELECT
platform,
COUNT(DISTINCT user_id) AS users,
ROUND(COUNT(*) FILTER (WHERE event_type = 'habit_logged') * 1.0 / COUNT(DISTINCT user_id), 1) AS habits_per_user
FROM input
GROUP BY platform
ORDER BY habits_per_user DESCRows
| Platform | Users | Habits per user |
|---|---|---|
| android | 26 | 19.7 |
| ios | 26 | 14.3 |
| web | 8 | 5.6 |
Step 2
Every iOS and web row is a distinct event. Android has hundreds of repeats: the client retries uploads on a bad connection and the pipeline keeps both copies.
| Platform | Rows | Events | Duplicate rows |
|---|---|---|---|
| android | 995 | 688 | 307 |
| ios | 736 | 736 | 0 |
| web | 134 | 134 | 0 |
Tableau
COUNT([event_id]) - COUNTD([event_id])The xmetric engine ran this formula and got the SQL’s Duplicate rows column, row for row.
SQL
SELECT
platform,
COUNT(*) AS ROWS,
COUNT(DISTINCT event_id) AS events,
COUNT(*) - COUNT(DISTINCT event_id) AS duplicate_rows
FROM input
GROUP BY platform
ORDER BY duplicate_rows DESCRows
| Platform | Rows | Events | Duplicate rows |
|---|---|---|---|
| android | 995 | 688 | 307 |
| ios | 736 | 736 | 0 |
| web | 134 | 134 | 0 |
Step 3
Counted by distinct event id, Android drops just behind iOS. The engagement gap was a retry bug.
Tableau
COUNTD(IF [event_type] = "habit_logged" THEN [event_id] END) / COUNTD([user_id])SQL
SELECT
platform,
ROUND(COUNT(DISTINCT CASE WHEN event_type = 'habit_logged' THEN event_id END) * 1.0 / COUNT(DISTINCT user_id), 1) AS habits_per_user
FROM input
GROUP BY platform
ORDER BY habits_per_user DESCRows
| Platform | Habits per user |
|---|---|
| ios | 14.3 |
| android | 13.7 |
| web | 5.6 |
That's the case for a visual first pass: one drag of event_id onto the view as a count and a distinct count shows the gap before anyone builds a dashboard on it.
Common mistake
Each week counts its own distinct users, so a user active in all four weeks is counted four times. Count distinct users over the whole period instead.
Tableau
WINDOW_SUM(COUNTD([user_id]))
// view at week levelSQL
WITH w AS (
SELECT
date_trunc('week', event_date) AS week,
COUNT(DISTINCT user_id) AS users
FROM input
GROUP BY week
)
SELECT SUM(users) AS active_users
FROM wTableau
COUNTD([user_id])
// view at month level, or { FIXED : COUNTD([user_id]) }SQL
SELECT COUNT(DISTINCT user_id) AS active_users
FROM inputDrop a CSV export of your own table and the same steps run on it, in this browser. For Parquet, open the file in the xmetric app; nothing is uploaded either way.
Drop a CSV anywhere on this page
Up to 10,000 rows. The sample, events.csv, shows the shape.
If this is you
A Parquet or CSV file on your laptop and a question to answer today
xmetric
Open it in a tab and drag fields; no server, no repo, no upload.
If this is you
A recurring report that should be reviewed in pull requests and deployed by CI
Evidence
SQL and Markdown in Git, built to a static site. Its hosted Team plan costs $2,500 / month, unlimited users, monthly billing; Core is free to self-host.
If this is you
A custom data app with hand-built charts in JavaScript
Observable Framework
Data loaders in any language precompute snapshots at build time; you write the charts.
If this is you
A team portal on a warehouse with roles and a shared SQL editor
Apache Superset / Preset
Free to self-host; Preset hosts it, with Professional at $20/user/month billed annually; $25/user/month billed monthly.
No. Parquet, CSV and Excel files are read inside your browser tab. Nothing leaves your machine unless you choose to save the workbook online.
No. The xmetric app opens .parquet files directly. The demo on this page uses a CSV copy of the same table so it loads instantly here.
Yes. Every step shows DuckDB SQL over a table called input. Replace input with read_parquet('events.parquet') and it runs as is.
Because the same event can arrive twice. Pipelines that deliver at least once, and mobile clients that retry uploads, write duplicate rows with the same id. COUNT counts rows; COUNTD counts events.