Wednesday
Hi all,
On an Azure Databricks workspace (Premium, Southeast Asia) that uses Lakebase, Predictive Optimization started running on 16 Nov 2026 on the Databricks-managed telemetry tables in system.lakebase (postgres_logs, compute_counters, plan_history and 7 others). It runs VACUUM, DELETE and PURGE every day and is billed as billing_origin_product = PREDICTIVE_OPTIMIZATION (SKU PREMIUM_JOBS_SERVERLESS_COMPUTE). We have no tables of our own that use it.
What we tried:
1. Disabled Predictive Optimization at account level (Account Console > Settings > Feature enablement). The metastore and the system catalog now show DISABLE.
2. But the schema system.lakebase still has enable_predictive_optimization = ENABLE set explicitly, and its tables inherit it. Operations kept running after the change (per system.storage.predictive_optimization_operations_history).
3. ALTER SCHEMA system.lakebase DISABLE PREDICTIVE OPTIMIZATION and the Unity Catalog API (schemas update) both fail with: PERMISSION_DENIED: Predictive optimization can't be modified on a system schema.
Usage so far: 2.58, 4.08 and 1.72 DBU per day. Small, but for a small R&D account it is the largest item on the bill.
Questions:
- Is there a supported way to turn Predictive Optimization off for system.lakebase, or to make it honour the account-level setting?
- Is this maintenance of system telemetry tables expected to be billed to the customer?
Thanks!
Wednesday
Hi,
Based on the Databricks documentation, disabling Predictive Optimization at the account level does not override an explicit ENABLE setting at the catalog or schema level. This explains why operations may continue on system.lakebase even after disabling it at the account level.
Since system.lakebase is a Databricks-managed system schema and both the SQL command and Unity Catalog API return PREMISSION_DENIED, there doesn't appear to be a documented self-service way to disable Predictive Optimization specifically for this schema.
Regarding billing, Databricks documents Predictive Optimization as billable serverless compute, but the documentation doesn't explicitly clarify whether maintenance of Databricks-managed Lakebase telemetry tables should be charged to customers.
I'd recommend raising a Databricks Support ticket to confirm whether this is expected behavior or an unintended configuration, and whether the charges can be reviewed.
References:
Do let us know once you receive a response from Databricks Support, as it would be helpful for others facing similar issues.
Thursday
Hi @trantu Predictive optimization uses an inheritance model. When enabled for a catalog, schemas in that catalog inherit the setting, and tables within an enabled schema inherit it as well. You can explicitly enable or disable predictive optimization for a catalog, schema, or table to override this behavior. Br
Saurabh
Thursday
Thanks @Divya_Pathak and @saurabh18cs.
Following the inheritance hint, I also tried disabling Predictive Optimization at table level (UC API, enable_predictive_optimization = DISABLE on system.lakebase.postgres_logs). It is refused too: "System owned table can't be modified. Table 'system.lakebase.postgres_logs' is a system owned table."
So every level we can control is blocked: account (DISABLE, but not inherited), schema (PERMISSION_DENIED on a system schema) and table (system owned). There is no self-service way to stop it.
We have opened a billing support ticket with Azure for Azure Databricks, asking them to disable Predictive Optimization on system.lakebase for our account and to review the charges since 16 Nov. I will post the outcome here once they reply.
Thursday
Hi @trantu,the table-level test helps narrow this down. The account-level setting supplies a default. As @Divya_Pathak pointed out, an explicit schema-level ENABLE can override it. Your schema and table requests then hit a separate restriction on system-owned objects. Given those results, I canโt identify a documented customer-side override for system.lakebase. Iโd ask Azure Databricks Support to confirm whether the service can change that setting for your account.
The billing question needs its own evidence. Azure Databricks documents Predictive Optimization as using billable serverless jobs compute. It also documents that the DBUs shown against individual operations are estimates when operations share compute. That explains how the numbers are recorded. Whether customers are intended to pay for this specific Lakebase telemetry maintenance still requires Databricks to confirm.
The relationship to verify is the explicit setting at the schema and the inherited setting on its tables. The two system tables then answer different questions. One records what ran, and the other records usage attributed to the workspace.
In a SQL session with read access to both system tables, set the session time zone to UTC. Replace the workspace ID and date with your affected workspace and the exact disable timestamp. Use the first query to tell whether work started afterward, and the second to capture overlapping billing records, including any usage corrections.
SET TIME ZONE 'UTC';
SELECT workspace_id, start_time, end_time, table_name,
operation_type, operation_id, operation_status,
usage_quantity AS estimated_dbu
FROM system.storage.predictive_optimization_operations_history
WHERE workspace_id = '1234567890123456'
AND catalog_name = 'system'
AND schema_name = 'lakebase'
AND start_time >= TIMESTAMP '2026-11-19 12:00:00'
ORDER BY start_time, table_name
;
SELECT workspace_id, usage_date, sku_name, usage_unit,
MIN(usage_start_time) AS first_usage_start_utc,
MAX(usage_end_time) AS last_usage_end_utc,
SUM(usage_quantity) AS net_usage
FROM system.billing.usage
WHERE workspace_id = '1234567890123456'
AND billing_origin_product = 'PREDICTIVE_OPTIMIZATION'
AND usage_end_time > TIMESTAMP '2026-11-19 12:00:00'
GROUP BY workspace_id, usage_date, sku_name, usage_unit
ORDER BY usage_date, sku_name
;For example, a VACUUM on postgres_logs that started at 13:05 UTC after a 12:00 UTC disable is evidence that another operation started. A positive billing record for the same workspace and window is evidence of Predictive Optimization usage. Treat that comparison as correlation. A billing interval can overlap the disable, and the workspace total can include other operations. If you see billed usage but no fresh Lakebase operations, widen the operation-history window to check work already in flight and check for other schemas in that workspace.
For the case, send the UTC setting-change time, the schema and table rejection messages, a few post-change operation IDs, and the corresponding SKU and usage windows. Ask for two explicit answers. Is there a supported opt-out for this system schema? Are these system.lakebase operations expected to be customer-billed, and can the usage be reviewed if they are not? If Support changes the configuration, success is no new operations starting after the agreed effective time and no new usage attributable to them once billing records settle.
If you can share whether the latest operation started after the disable, that would distinguish recurring activity from usage that arrived later. A kudo is appreciated if the queries help you document the case.
References
yesterday
Short answer: what you're seeing is how the inheritance model works. The account-level setting is only a default, and an explicit schema-level setting overrides it. Because system.lakebase is a Databricks-managed schema, you can't change that override yourself, so I don't know of a customer-side toggle that fixes this. You'll need a support ticket.
Predictive Optimization (PO) is inherited from account to catalog to schema to table, and you can override it at any level. Disabling it at the account level explicitly does not disable it for catalogs or schemas that have specifically enabled it [1]. system.lakebase has ENABLE set explicitly on the schema, so its tables follow that value regardless of what the metastore or the system catalog shows. That matches your observation that operations kept running after you changed the account setting.
The documented way to change PO on a schema is ALTER SCHEMA ... { ENABLE | DISABLE | INHERIT } PREDICTIVE OPTIMIZATION, which normally requires being the schema owner or having MANAGE on it [1]. For system.lakebase you don't own the schema, and the platform rejects the change with PERMISSION_DENIED: Predictive optimization can't be modified on a system schema. The SQL and the UC API hit the same guard, so no alternative path will get around it. I wouldn't try to work around it with grants either, because the block is on the schema type, not on your privileges.
PO runs OPTIMIZE, VACUUM, ANALYZE and REFRESH INDEX on serverless compute for jobs, and the account is billed for it under a serverless jobs SKU [1]. That is consistent with the PREDICTIVE_OPTIMIZATION / PREMIUM_JOBS_SERVERLESS_COMPUTE line you're seeing. What the product description doesn't spell out is whether maintenance on Databricks-managed telemetry tables is meant to be charged to the customer. I wouldn't assume it is intended, particularly for a schema you can't opt out of. It's a fair question for Databricks and Microsoft, and you have a clear case.
Open a support ticket through the Azure portal (Microsoft support for Azure Databricks) or your Databricks account team. Include:
enable_predictive_optimization = ENABLE on system.lakebasePERMISSION_DENIED errorAsk two things: (a) can PO on system.lakebase be disabled or set to inherit from the account, and (b) can the usage be credited or excluded from billing if it's platform-initiated maintenance.
Keep the evidence. These two queries give you the numbers for the ticket:
SELECT usage_date, SUM(usage_quantity) AS dbus
FROM system.billing.usage
WHERE billing_origin_product = 'PREDICTIVE_OPTIMIZATION'
GROUP BY usage_date
ORDER BY usage_date;
SELECT *
FROM system.storage.predictive_optimization_operations_history
WHERE catalog_name = 'system' AND schema_name = 'lakebase'
ORDER BY start_time DESC;
Keep the account-level setting as it is. Leaving it disabled is still the right default for your own catalogs, since it stops PO from being applied to new objects through inheritance. It just doesn't touch schemas with an explicit ENABLE.
At 2โ4 DBU/day this is small, but it's worth pushing on, because a setting that a customer can't change and that also generates a bill is the kind of thing support can escalate.
[1] Predictive optimization for Unity Catalog managed tables - Azure Databricks | Microsoft Learn โ https://learn.microsoft.com/en-us/azure/databricks/optimizations/predictive-optimization