SQL Execution
Error E-3113 on long string columns in 2.1.x
Symptom
SELECT statements fail with [E-3113] string column length after upgrading to 2.1.0 or 2.1.1.
Fix
Raise the session variable parallel_pipeline_task_num. On a 32-core / 256 GB node:
SET parallel_pipeline_task_num = 16; -- or 32
JDBC catalog connection exhaustion
Symptom
Can not connect to jdbc due to error: wait millis 5000, active 10, maxActive 10, creating 0
Fix
Increase the pool size or recreate the catalog:
ALTER CATALOG <catalog_name>
SET PROPERTIES ('connection_pool_max_size' = '100');
Data Definition
varchar vs. string
| Use case | Recommendation |
|---|---|
| Key column | Use VARCHAR; STRING is not allowed |
| Non-key column ≤ 65 533 chars | VARCHAR(65533) |
| Non-key column > 65 533 chars | STRING |
VARCHAR(65533)reserves at most 65 533 UTF-8 code points; actual storage equals the real data length.
Multi-statement jobs
CREATE JOB in 2.1 supports only one SQL statement per job. A DAG-style task orchestration is on the roadmap.
Asynchronous materialized views
Refresh granularity is currently minute-level. Sub-second refresh is planned via synchronous multi-table materialized views.
Arrow Flight port
The Arrow Flight SQL port can differ from the FE port; any free port is acceptable.
DECIMAL versions after 1.2 → 2.0 upgrade
New tables always use DECIMAL V3. Legacy tables retain their original type; alter the column manually to migrate.
Administration
information_schema restrictions
information_schema exposes virtual views only—no physical files exist—so DROP, ALTER, or INSERT are rejected.
priority_networks with VIP (keepalived)
Given:
- Physical IP range:
10.0.0.0/24 - VIP:
10.0.0.10(managed by nginx+keepalived) - Desired subnet mask:
10.0.0.128/25
FE selects the first IP matching priority_networks. If no IP in 10.0.0.128/25 is present, the FE exits. Ensure at least one server IP falls inside the new subnet; otherwise adjust the mask or add a interface.
Replication
Raising the replica count from 3 to 5 on a 5-BE cluster increases fault tolerance (survives two simultaneous BE failures) but does not linearly improve query speed. Use extra replicas primarily for availability, not performance.