#7211 Fix non-deterministic workflow digest and oversized error messages Pull request opened by andresgomezfrr Tracking issue Closes #7212 Related to #4780 Why are the changes needed? When re-registering a workflow with the same version, FlyteAdmin computes a digest of the compiled workflow to check if the structure is identical. Two bugs cause this to fail for large workflows: Bug 1: Non-deterministic digest (compiler)

ValidateWorkflow

in

workflow_compiler.go

iterates

wf.Nodes

(a Go map) in two places without sorting keys. Since Go map iteration is randomized, the same workflow produces different

CompiledWorkflowClosure

outputs across compilations, leading to different digests. FlyteAdmin then incorrectly takes the "different structure" code path instead of returning

ALREADY_EXISTS

. Bug 2: Oversized error message (errors) The "different structure" code path computes a

jsondiff

of two large compiled workflow closures and includes the full diff in the gRPC error description. For workflows with many dependencies (e.g. 400+ JARs), this diff exceeds gRPC's default 4MB

MaxSendMsgSize

. gRPC-Go silently rejects the response at the transport layer, sending

RST_STREAM INTERNAL_ERROR

to the client with no server-side log. The client sees

INTERNAL: RST_STREAM closed stream. HTTP/2 error code: INTERNAL_ERROR

instead of any useful error. What changes were proposed in this pull request? Commit 1: Truncate diff in error messages (

flyteadmin/pkg/errors/errors.go

) • Cap error messages at 3MB (leaving room for gRPC framing) in all three

*ExistsDifferentStructureError

functions • When truncated, append

... [diff truncated — exceeded gRPC max message size]

• Safety net that prevents RST_STREAM regardless of diff size Commit 2: Sort node IDs for deterministic compilation (

flytepropeller/pkg/compiler/workflow_compiler.go

) • Sort

wf.Nodes

map keys before iterating in

ValidateWorkflow

• Ensures identical workflows produce identical

CompiledWorkflowClosure

and therefore identical digests • Root cause fix: re-registration of identical workflows correctly returns

ALREADY_EXISTS

How did you test it? • All existing tests pass:

go test ./pkg/errors/

and

go test ./pkg/compiler/...

• Reproduced the original issue on a production FlyteAdmin instance by calling

CreateWorkflow

with a modified template (same version) — confirmed

RST_STREAM INTERNAL_ERROR

• Same call with identical template returns

ALREADY_EXISTS

(correct behavior) •

CreateTask

always returns

ALREADY_EXISTS

(task digests are already deterministic — no map iteration in task compilation) flyteorg/flyte