[AURORA-1494] Scheduler fails to start due to thrift/SQL schema data type mismatch

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Fixed
• Fix Version/s: 0.10.0
• Component/s: Scheduler
• Sprint: Twitter Aurora Q3'15 Sprint 12
• Story Points: 3
• Assignee: Maxim Khutornenko
• Reporter: Maxim Khutornenko
• Created: 18/Sep/15 00:23
• Updated: 10/Nov/15 17:59
• Resolved: 18/Sep/15 17:13

Description

After https://reviews.apache.org/r/38288 we are unable to upgrade scheduler in one of our clusters due to the following failure on restart:

### Cause: org.h2.jdbc.JdbcSQLException: Numeric value out of range: "3174400000031744"; SQL statement:
INSERT INTO task_configs (
      job_key_id,
      creator_user,
      service,
      num_cpus,
      ram_mb,
      disk_mb,
      priority,
      max_task_failures,
      production,
      contact_email,
      executor_name,
      executor_data,
      tier
    ) VALUES (
      (
        SELECT ID
        FROM job_keys
        WHERE role = ?
          AND environment = ?
          AND name = ?
      ),
      ?,
      ?,
      ?,
      ?,
      ?,
      ?,
      ?,
      ?,
      ?,

This appears due to type mismatch between TaskConfig.diskMb (i64) and task_configs.disk_mb (INT).

A possible real-life scenario:

• user creates a job with an oversized resource requirement and the job fails to schedule
• user realizes the mistake and attempts to correct it by running aurora update start
• scheduler creates an instance of the JobUpdate with the oversized TaskConfig as its initial state and persists it in the log
• scheduler restarts to a new version (with the patch above) and attempts to reload job updates from the log but now instead of storing TaskConfigs as binary blobs it attempts to insert into task_configs table where resource columns have narrower type.

Issue Links

• blocks: AURORA-1495 Consider changing TaskConfig resource types to reflect reasonable values