Compute instances¶
Data processing, such as embedding and enrichment, SQL queries, and operator execution all require compute capacity. A compute instance is the workspace object used to manage that capacity. It has a specification, supports elastic scaling and automatic suspension, provides compute for workflows, the SQL editor, data loading/export, operator services, knowledge-base conversational queries, and other workloads, and is billed by node online time.
Entry point: Resource Center → Compute Instances.
Choose a type first: query or task¶
You must choose a type first when creating a resource. The type determines the engine and specification tiers, and the two types cannot substitute for each other.
Query compute instance |
Task compute instance |
|
|---|---|---|
Execution content |
SQL only |
Non-SQL processing only |
Typical uses |
SQL editor, workflow SQL nodes, structured reads/writes on MO tables, knowledge-base NL2SQL |
Non-SQL workflow operators, document parsing/embedding, load and export processing, operator APIs |
Specifications |
Query compute tiers |
General-purpose CPU / GPU compute |
How to choose¶
SQL reads or writes that touch MO tables → query resource
Non-SQL processing → task resource
If a workflow or structured load/export uses both sides → attach one of each type
Consumer |
Query resource |
Task resource |
|---|---|---|
SQL editor (MOI tables / federated queries) |
✓ |
— |
Workflow |
SQL/table read-write nodes |
Non-SQL nodes |
Structured load/export |
Read/write MO tables |
Fetch/parse/map/external write |
Unstructured load/export |
— |
✓ |
Operator API |
— |
✓ |
Knowledge-base conversational query (NL2SQL) |
✓ |
— |
Choosing a specification economically¶
Start with a small tier and scale up only when needed.
Choose a GPU tier only when a GPU is actually required; CPU is sufficient for parsing and routine processing.
For memory-intensive tasks, prioritize a higher memory ratio instead of blindly moving up an entire tier.
For intermittent workloads, prefer a small specification with elasticity; consider always-on resources for latency-sensitive scenarios.
For list prices, see Credit and billable items.
Default resources and billing¶
The platform provides one default resource for each of the query and task types. They are marked “Default” in the list and cannot be deleted. When no resource is specified elsewhere, the corresponding default resource is used automatically. It can generally scale to 0 while idle and start again when a task arrives.
Rule |
Description |
|---|---|
Billing unit |
Credit (standard price is approximately 10 Credit = CNY 1; top-up discounts may apply) |
Billing basis |
Ready online duration: exact active replica seconds across 1-minute (UTC) windows while nodes are Ready, converted to hours ( |
Zero charge windows |
Starting / Provisioning, Stopping, Failed, and Suspended (0 nodes, manual or auto-suspend) states do not generate running charges |
Pre-activation admission |
Account Credit sufficiency is verified before activating or adding compute workloads; sample exemption workspaces skip compute deductions |
SQL |
Not billed separately; SQL workloads share query compute instances |
For complete compute and sandbox billing rules, see How compute instances and sandboxes are billed.
The Credit badge in the top bar shows your balance. The resource list and details show current hourly consumption. For bill analysis, see View cost and usage.
Elastic scaling and automatic suspension¶
Setting |
Purpose |
|---|---|
Minimum/maximum node count |
Automatically scale within this range according to load |
Automatic suspension |
Scale to 0 after N idle minutes, such as 5, 10, or 30 minutes or 1 hour, or never suspend automatically |
Request wake-up |
A new request automatically starts a suspended resource; the first request may encounter a cold start |
The same settings cover two operating patterns:
Pattern |
Configuration approach |
Suitable for |
|---|---|---|
Always-on / low latency |
Minimum = maximum = 1; no automatic suspension |
Online APIs and latency-sensitive workloads |
Elastic / cost-saving |
Set a node range and enable automatic suspension |
Batch and intermittent workloads |
Create a compute instance¶
Go to Resource Center → Compute Instances and click Create Compute Resource.
Enter a name (required) and description (optional).
Select a type: query or task.
Select a specification. Its unit price and configuration are shown on the card.
Configure the minimum and maximum node counts, such as the default 1–2.
Configure the automatic suspension policy, such as the default 10 minutes.
Create the resource. It transitions from “Starting” to “Running”.
List and routine management¶
You can search the list and filter by type or status. Common columns include name, type, specification, node count (current/maximum), status, Credit consumption, associated workloads, automatic suspension, and creation time.
Operation |
Description |
|---|---|
Suspend |
Scale nodes to 0 and stop billing; associated workloads are interrupted |
Start |
Start manually; a new request can also wake a suspended resource |
Delete |
Irreversible; deleting a running resource interrupts workloads; default resources cannot be deleted |
Click a name to view details, including basic information, utilization, associated workloads, specification, and unit price.
Caution: Before suspending or deleting a resource, check its details to ensure no critical task is running.
Configure custom Python dependencies¶
When enabled on your deployment, Task compute instances with a Python worker support custom third-party Python packages via requirements.txt for custom Python workflow operators. Query compute instances (SQL) do not support Python dependencies. If custom dependencies are disabled on the platform or unsupported by the backend, the editor is disabled and displays the reason.
In the create or edit dialog for a task compute instance, locate the Python dependencies section.
Click Upload requirements.txt to load a local
.txtfile, or paste/edit dependencies directly in the text area.Check the byte and line usage counters below the editor and save once validation passes.
Follow these formatting and limit rules when editing requirements.txt:
Allowed entries: One PEP 508 requirement per line (package name, version specifier, extras), or
--index-url/--extra-index-urlpointing to a host in the platform’s allowed index list.Disallowed entries: Local filesystem paths (
/,./,../),file://URLs, VCS URLs (git+, etc.), and index hosts outside the allowlist are rejected.Size and line limits: Default limits are 262,144 bytes (256 KiB) and 200 non-empty lines; check the exact threshold shown beneath the editor for your deployment.
After saving dependency updates, open the instance detail view to track installation progress:
Overall Status (Ready, Installing, Failed, or No observation), ready instance ratio, content
SHA, update time, and the savedrequirements.txtsource;Last error and the Per-instance observation table showing status, SHA, and errors for each worker pod. To remove custom dependencies, clear the editor in the edit dialog and save.