Troubleshooting

Projects and storage

"This project is open in another tab"

Only one tab may edit a project. Close the other tab, or press Work here instead in the blocked one. See Projects.

I cannot save a project to a folder

Saving to a folder needs a Chromium browser (Chrome, Edge, Brave and similar). Firefox and Safari keep projects in the browser's own storage, which is fully supported but is not a folder you can open in a file manager.

Storage is full or slow

Open the Data Hub's Local Storage view. Remove files no dataset references, delete data you no longer need, or run the project cleanup, which shows what it will remove first. See Profiles, versions and lineage.

Importing data

A file imported but shows as binary or will not preview

Some formats are stored but not read: Python pickles (.pkl, .pt, .pth, .joblib) are never opened because loading them can run code, and ONNX models are previewed rather than loaded. Convert to SafeTensors, NumPy or Parquet.

A large file fails or is cut off

Formats DuckDB cannot read natively are limited to 16 MB or 200,000 rows. Convert to Parquet or CSV. Large CSVs use a streaming parser that needs UTF-8, a header row and double quotes. See Importing data.

The same file keeps being flagged as a duplicate

DLWAY compares file content. Choose Keep copy to import it again, or Link it to point a new dataset at the existing bytes.

A database table looks incomplete

Live previews show only the first rows, and snapshots are capped at 25,000 rows and say when they were truncated. Transform refuses a live preview; import a snapshot first. See Database connections.

A Kaggle or Hugging Face import fails

For Kaggle, reconnect your account (the last check's result is shown), and make sure you are signed in to DLWAY. For Hugging Face, a gated dataset needs its terms accepted on huggingface.co and a valid hf_ token; the message says which of the two is missing.

Remote compute

• "The kernel is busy": the kernel runs one request at a time. Wait for the current run, or press Stop.
• A connection to a local server is refused: the server must allow the DLWAY origin. Public servers go through the relay, which refuses private addresses.
• A run stopped after about two minutes: streams last up to 120 seconds (290 through the relay). Use the provider's own jobs for long training.
• "Too many changes to sync": a run uploads at most 300 files or 4 MB of changed text. Turn sync off, or reduce the number of changed files.
• A PrepFlow run on the kernel is refused: image steps that read browser-only media cannot run remotely. Run in the browser, or materialise the images first.

See Remote compute.

Training

Training is slow

Training uses WebGPU where available and WebGL otherwise. Check Browser requirements, close other GPU-heavy tabs and lower the batch size. The Model Builder's Resources tab shows the memory the run needs.

The Model Builder warns about memory before training

It compares the estimate with a budget. The browser cannot report GPU memory, so the budget starts as an assumption; set your GPU's real figure in Resources. Apply & train makes the suggested changes, Train anyway leaves the model alone.

The browser runtime is unavailable for my graph

Graphs that use Hugging Face blocks produce a Python script instead of running in the browser. Open the Code tab to take the script and notebook. See Visual Model Builder.

Pipelines

• A scheduled pipeline did not run: schedules run only while the Studio is open in a tab. A missed run either catches up once or is skipped, as set under Settings → Data and pipelines.
• A step says "branch not taken" or was skipped: a Branch, New Rows with nothing new, or an earlier stop left it with nothing to do. This is not a failure.
• A run ended "gated": the Quality Gate rejected the model, so packaging was withheld.
• New Rows finds nothing on the second run: a successful run moves its position forward, so running again straight away finds no new data. This also affects parameter sweeps.
• A write to a database waits for approval: that connection's write mode is Ask, Type to confirm or Dry run first. See Database connections.

See Running pipelines.

Dashboard

• PNG export says a chart did not respond: wait for every chart to finish drawing, then export again.
• A chart is empty after a dataset changed: check the dataset version under the toolbar, and whether a calculated field refers to a column that was renamed.

AI Copilot

• A tool call never appears: in Ask mode the agent waits for your approval. Look for the approval card.
• An MCP server cannot connect: only HTTP servers that speak SSE work in a browser, and the server must allow the Studio's origin.
• A hook did not run a linter: hooks cannot run command-line tools in a browser; they show their text instead. Use a Python hook or the diagnostics hook.

Still stuck?

Use the contact page and tell us the module, what you expected and what happened.