CLI version: v1.14.0 (also present on v1.13.0)
Engine: direct
Summary
The CLI sends a fixed update_mask on every UpdateApp call, independent of what the bundle declares or what actually changed. compute_size is always in it. But the SDK field is compute_size,omitempty, so when the bundle does not declare a compute size the field is absent from the request body — the request asks the backend to update a field it does not supply.
Evidence
Both halves are captured requests, not inference. Taken with --debug and DATABRICKS_DEBUG_TRUNCATE_BYTES=20000 (the default truncation hides the tail where update_mask lives).
With compute_size: MEDIUM declared — value present, mask lists it:
{
"app": {
"compute_size": "MEDIUM",
"description": "...",
"name": "<app>",
"resources": [ ... ],
"source_code_path": ""
},
"update_mask": "description,budget_policy_id,usage_policy_id,resources,user_api_scopes,forward_user_access_token,compute_size,compute_min_instances,compute_max_instances,git_repository,telemetry_export_destinations"
}
With the declaration removed — same bundle, same target, only that line deleted. The body now has no compute_size key at all, and the mask is byte-identical, still naming it:
{
"app": {
"description": "...",
"name": "<app>",
"resources": [ ... ],
"source_code_path": ""
},
"update_mask": "description,budget_policy_id,usage_policy_id,resources,user_api_scopes,forward_user_access_token,compute_size,compute_min_instances,compute_max_instances,git_repository,telemetry_export_destinations"
}
So the mask is a fixed list rather than being derived from what the request actually carries.
The direct engine's own embedded resources.generated.yml already classifies the field as a backend default:
apps:
backend_defaults:
# Backend sets it "MEDIUM" when not specified in the config
- field: compute_size
So the CLI knows compute_size is a backend default for diffing, and still masks it for update. compute_min_instances and compute_max_instances are in the mask on the same terms and are worth checking alongside it.
Why plan cannot show this
bundle plan correctly reports compute_size as {"action":"skip","reason":"backend_default"}. The differ is not what is inconsistent — the update payload is, and plan never renders a payload. Any bug of this shape is invisible to the documented pre-deploy gate.
Suggested fix
Build the update_mask from the fields actually present in the serialised body, or drop backend_defaults fields from the mask when the config does not declare them — the same treatment PR #5042 gave deploy-only fields.
What we are NOT claiming
We have not observed the backend mishandling the masked-but-absent field. The undeclared payload above was captured on v1.14.0, where the request is rejected with a 400 on an unrelated field (source_code_path, reported separately) before the backend acts on the mask — so what this service does with "update compute_size" plus no value is untested here.
Nor is this the explanation for the Unexpectedly failed to update app's compute size errors, which is where the tempting story leads. Three consecutive deploys on 2026-08-27 with compute_size declared and present in the body produced a Lakebase grant timeout (deadline exceeded after 998000000ns), then that compute-size message, then a clean success — same config, minutes apart. That reads as a flaky async update path, not as this payload bug, and an earlier internal write-up that claimed otherwise was withdrawn on this evidence.
This report is about payload correctness only: the mask names a field the request does not carry.
CLI version: v1.14.0 (also present on v1.13.0)
Engine:
directSummary
The CLI sends a fixed
update_maskon everyUpdateAppcall, independent of what the bundle declares or what actually changed.compute_sizeis always in it. But the SDK field iscompute_size,omitempty, so when the bundle does not declare a compute size the field is absent from the request body — the request asks the backend to update a field it does not supply.Evidence
Both halves are captured requests, not inference. Taken with
--debugandDATABRICKS_DEBUG_TRUNCATE_BYTES=20000(the default truncation hides the tail whereupdate_masklives).With
compute_size: MEDIUMdeclared — value present, mask lists it:{ "app": { "compute_size": "MEDIUM", "description": "...", "name": "<app>", "resources": [ ... ], "source_code_path": "" }, "update_mask": "description,budget_policy_id,usage_policy_id,resources,user_api_scopes,forward_user_access_token,compute_size,compute_min_instances,compute_max_instances,git_repository,telemetry_export_destinations" }With the declaration removed — same bundle, same target, only that line deleted. The body now has no
compute_sizekey at all, and the mask is byte-identical, still naming it:{ "app": { "description": "...", "name": "<app>", "resources": [ ... ], "source_code_path": "" }, "update_mask": "description,budget_policy_id,usage_policy_id,resources,user_api_scopes,forward_user_access_token,compute_size,compute_min_instances,compute_max_instances,git_repository,telemetry_export_destinations" }So the mask is a fixed list rather than being derived from what the request actually carries.
The direct engine's own embedded
resources.generated.ymlalready classifies the field as a backend default:So the CLI knows
compute_sizeis a backend default for diffing, and still masks it for update.compute_min_instancesandcompute_max_instancesare in the mask on the same terms and are worth checking alongside it.Why
plancannot show thisbundle plancorrectly reportscompute_sizeas{"action":"skip","reason":"backend_default"}. The differ is not what is inconsistent — the update payload is, andplannever renders a payload. Any bug of this shape is invisible to the documented pre-deploy gate.Suggested fix
Build the
update_maskfrom the fields actually present in the serialised body, or dropbackend_defaultsfields from the mask when the config does not declare them — the same treatment PR #5042 gave deploy-only fields.What we are NOT claiming
We have not observed the backend mishandling the masked-but-absent field. The undeclared payload above was captured on v1.14.0, where the request is rejected with a 400 on an unrelated field (
source_code_path, reported separately) before the backend acts on the mask — so what this service does with "updatecompute_size" plus no value is untested here.Nor is this the explanation for the
Unexpectedly failed to update app's compute sizeerrors, which is where the tempting story leads. Three consecutive deploys on 2026-08-27 withcompute_sizedeclared and present in the body produced a Lakebase grant timeout (deadline exceeded after 998000000ns), then that compute-size message, then a clean success — same config, minutes apart. That reads as a flaky async update path, not as this payload bug, and an earlier internal write-up that claimed otherwise was withdrawn on this evidence.This report is about payload correctness only: the mask names a field the request does not carry.