executor crate architecture

noetl-executor workspace crate

Since v2.17.1+ the CLI is a Cargo workspace. The root crate (noetl, producing the noetl and ntl binaries) is unchanged; a new workspace-member crate noetl-executor hosts the shared execution utilities and types that both this CLI and the noetl-worker daemon (R-1.3) depend on.

This page documents the executor crate's module layout, what moved out of playbook_runner.rs and why, and the deliberate architectural decision that the CLI keeps its own control loop rather than sharing one with the worker.

Source layout

repos/cli/
  Cargo.toml                    # workspace root + noetl binary
  src/
    main.rs
    playbook_runner.rs          # CLI tree walker (control loop stays here)
    ...
  executor/                     # workspace member: noetl-executor
    Cargo.toml
    src/
      lib.rs
      playbook.rs               # YAML playbook types
      template.rs               # render_template + Rhai/JSON helpers
      condition.rs              # evaluate_condition + evaluate_rhai_condition
      capabilities.rs           # validate_capabilities + ValidationReport
      runtime.rs                # ExecutionContext + CredentialResolver trait
      events.rs                 # EventSink trait + ExecutorEvent + EventEmitter
      tools_bridge.rs           # noetl-tools registry bridge (scaffold)
      worker/
        mod.rs
        source.rs               # Command + CommandSource (worker-only)

What lives in noetl-executor and why

Module	Purpose	Used by
playbook	Pydantic-like YAML types: Playbook, Step, Tool, NextFormat, RuntimeCapabilities, etc. All field accessors pub.	CLI + worker
template	render_template, render_template_with_result, get_json_path, json_to_rhai, rhai_to_json_string. Takes &HashMap<String, String> views of the per-execution variables + step results so each binary owns its own context shape.	CLI + worker
condition	evaluate_condition (simple {{ a == b }} / 'in' / truthy) and evaluate_rhai_condition (full Rhai expression eval). Same context-view contract as template.	CLI + worker
capabilities	validate_capabilities + ValidationReport / ValidationError. Pure function: returns the report rather than bail!ing so the CLI can format errors with playbook_path and the worker with execution_id.	CLI + worker
runtime	ExecutionContext (executor-side variant with async step_results + Arc<dyn CredentialResolver>); CredentialResolver trait.	worker (CLI uses its own ExecutionContext in playbook_runner.rs)
events	ExecutorEvent (mirrors the Python noetl.runtime.events.report_event envelope), EventSink trait, NoopSink, EventEmitter.	CLI + worker
tools_bridge	Scaffold for replacing the CLI's inline tool implementations with calls into the noetl-tools registry. Filled in incrementally per Strategy B (one tool kind per sub-PR).	CLI (worker already uses noetl-tools directly)
worker::source	Command envelope + CommandSource trait. Worker-only: the CLI's tree walker doesn't consume this.	worker (R-1.3)

What does NOT live in noetl-executor

	Where it lives	Why
The CLI's recursive tree walker (run, execute_step, execute_next_steps, execute_router_arcs)	repos/cli/src/playbook_runner.rs	Natural fit for local YAML execution; flattening into a pull-model iterator would lose local-debug clarity
The CLI's inline tool implementations (execute_tool, execute_shell_command, execute_http_request, execute_duckdb_query, etc.)	repos/cli/src/playbook_runner.rs (today)	Migrated incrementally to noetl-tools registry via tools_bridge per Strategy B; ~870 LoC of inline tool dispatch
The worker's NATS pull loop	repos/worker/src/	Different shape than tree walker; tied to NATS durable-consumer semantics
RunOutcome (the CLI's JSON output envelope)	repos/cli/src/playbook_runner.rs	Not a YAML input type; worker has a different output envelope (event-log writes)

Architectural finding (§ H.10 of Appendix H)

Mid-implementation discovery, documented in § H.10 of the global hybrid cloud blueprint:

• The CLI is a recursive tree walker. It loads the YAML, walks the workflow, evaluates next arcs / case conditions / then blocks in place, dispatches each step inline. Control flow is the call stack.
• The worker is a pull-model consumer. It subscribes to a NATS durable consumer, pulls one command at a time, executes it, emits events, repeats. No tree. No recursion.
• The original "unified CommandSource trait — both binaries plug in their own impl" was the wrong abstraction for the CLI. Flattening the tree walker into a pull-model iterator loses local-debug clarity, complicates case / then state management, and breaks integration tests written against the tree shape.

noetl-executor was re-scoped from a control-loop crate to a utilities-and-types crate. The CLI's tree walker stays. The worker's pull loop stays. Both call into the executor for the same template rendering, condition evaluation, credential resolution, capability validation, and event shape.

Cross-references

Topic	Where
The migration roadmap	Appendix H of the global hybrid cloud blueprint
The architectural finding	§ H.10 of the same doc
The Polars-pattern endpoint (pip install noetl ships Rust runtime + Python wrapper)	§ H.9
Apache Arrow data plane	§ H.4 + § H.11 (local-mode Feather buffer)
Tracking issues	noetl/cli#19 (this CLI's R-1.1 sub-issue) · noetl/ai-meta#30 (umbrella)

Sub-PR landing history

Sub-PR	Scope	PR
R-1.1 PR-1	Crate skeleton (lib.rs, runtime, events, source/dispatch placeholders)	#20
R-1.1 PR-2a	YAML types → noetl-executor::playbook	#21
R-1.1 PR-2b	Utilities (template + condition + capabilities); § H.10 restructure (placeholder LocalPlaybookSource removed; CommandSource → worker::source)	#22
R-1.1 PR-2c-1	noetl-tools = "2.8.7" dep + tools_bridge scaffold	#23
R-1.1 PR-2c-2	tools_bridge adapters: BridgeContext, to_tools_context, to_tools_config (all 8 Tool variants), from_tools_result; dispatch_via_registry becomes async with per-tool-kind match arms	#24
R-1.1 PR-2c-3	Tool::Rhai arm wired through noetl-tools::RhaiTool; new to_tools_context_for_rhai helper groups CLI flat variables into nested Maps for Rhai field access; CLI's inline execute_rhai_script + rhai_http_request deleted (~220 LoC)	#25
R-1.1 PR-2c-4	Tool::Shell arm wired through noetl-tools::ShellTool; per-command dispatch loop preserves CLI's "fresh bash invocation per command" semantics; new shell_command_config(&str) helper; CLI's inline execute_shell_command deleted (~79 LoC)	#26
R-1.1 PR-2c-5	Tool::Http arm wired through noetl-tools::HttpTool; new resolve_auth_to_bearer helper (CLI AuthConfig → Bearer token via noetl-tools::auth::GcpAuth); new http_tool_config helper that injects the Bearer header into request headers; new reshape_http_result helper that maps noetl-tools' {status_code, headers, body} envelope back to the CLI's pre-existing {status, body} shape so playbook steps keep branching on <step>.body.status. Tool::Auth arm also routes through resolve_auth_to_bearer so both paths share the GCP ADC code path. CLI's inline execute_http_request + get_auth_token deleted (~148 LoC)	#27
R-1.1 PR-2c-6	Tool::DuckDb arm wired through noetl-tools::DuckdbTool; new duckdb_tool_config helper (translates CLI's Vec<String> params to noetl-tools' Vec<serde_json::Value> and maps db → db_path); new reshape_duckdb_result helper unwraps noetl-tools' {columns, rows, row_count} envelope back to the CLI's pre-existing rows-array shape (and {affected_rows} back to {"status": "ok"} for non-SELECT). Empty / whitespace-only query short-circuits to an empty outcome, mirroring the CLI's existing guard. Path resolution + mkdir -p stay at the CLI call site since the bridge has no knowledge of the playbook directory. CLI's inline execute_duckdb_query deleted (~55 LoC). Feature gain: params that were silently ignored (_params: &[String]) are now bound at ? placeholders	#28
R-1.1 PR-2c-7	Codifies the § H.10 finding for Tool::Playbook — sub-playbook execution stays in the CLI's tree walker (PlaybookRunner::new(path).run() is the recursion case; the bridge cannot replace it without re-opening § H.10). Replaces the silent BridgeOutcome::empty() stub with anyhow::bail! naming § H.10 so accidental misuse is loud rather than silent. Variable preparation (parent merge + DSL v2 input: / DSL v1 args: rendering + workload. prefix) DID move into a new noetl_executor::tools_bridge::prepare_sub_playbook_vars helper — pure, reusable, testable. CLI's call site uses the helper but keeps its PlaybookRunner recursion inline. No semantic divergences (no execution path changed)	#29
R-1.1 PR-2c-8	Final substantive sub-PR. Codifies that Tool::Auth and Tool::Sink stay inline by design but extracts the shareable pure logic into bridge helpers: auth_context_updates(provider, token, project) (replaces inline set_variable calls; preserves CLI's pre-PR-2c-8 ordering), format_sink_payload(format, raw) (json passthrough / yaml dump / csv conversion), and json_to_csv(json_str) (lifted verbatim from CLI). Dispatch arms for both kinds bail loudly with helper-pointing messages. CLI's inline json_to_csv deleted (~42 LoC). Removes the dispatch_via_registry_returns_empty_for_unwired_kind test — every Tool variant now either dispatches through the registry, bails with a § H.10 finding, or bails as unsupported. GCS → object_store migration tracked as a separate follow-up (R-2.x scope, not R-1.x). No semantic divergences	#30

After PR-2c-8, the remaining R-1.1 work is PR-2d: integration-test pass + final docs + closing noetl/cli#19.

Documented semantic divergences (per-PR cross-reference)

Each tool replacement that crosses a behaviour line documents the deltas in the PR body and here.

PR	Surface	CLI behaviour (pre-replacement)	noetl-tools behaviour	User-visible impact
#25	rhai timestamp()	chrono::Local::now().format("%H:%M:%S") → "14:23:45"	chrono::Utc::now().timestamp().to_string() → "1716847425"	Scripts that displayed timestamp() for human reading need updating
#25	rhai HTTP helpers (http_get, http_post, http_delete, *_auth)	curl subprocess	Direct reqwest calls	Same surface; different error shape on network failures
#25	rhai get_gcp_token	gcloud auth print-access-token shellout	gcp_auth crate (workload-identity aware)	Better on GKE pods; equivalent on hosts with gcloud configured
#26	shell stdout streaming	Line-by-line to terminal as command runs	Collected; returned at end	Breaks real-time output UX for long-running shell steps — users see nothing until command completes
#26	shell result envelope	Captured stdout string	data: {exit_code, stdout, stderr} JSON	Transparent — bridge unwraps data["stdout"] and trims trailing \n
#27	http transport	curl subprocess	reqwest direct	Same envelope on success; different error shape on transport failure (anyhow message vs curl exit code)
#27	http auth (GCP ADC)	gcloud auth print-access-token shellout	gcp_auth crate	Better on GKE pods (workload-identity aware); equivalent on hosts with gcloud configured
#27	http JSON body	Sent verbatim via curl -d; caller had to set Content-Type	Auto-detected as JSON; reqwest sets Content-Type	Most callers were already sending JSON with Content-Type: application/json; transparent
#27	http result envelope	{"status": int, "body": <json-or-string>}	data: {"status_code": int, "headers": {...}, "body": <json>}	Transparent — bridge's reshape_http_result maps back to the CLI's {status, body} shape so 4xx/5xx come back inside the envelope (not as anyhow errors)
#27	Tool::Auth GCP token	gcloud auth print-access-token shellout (separate code path from rhai get_gcp_token)	gcp_auth crate via shared resolve_auth_to_bearer helper	Both Tool::Http and Tool::Auth now share the GCP ADC code path — eliminates the divergence between the two
#28	duckdb SELECT result envelope	JSON array of row objects (pretty-printed)	data: {"columns": [...], "rows": [...], "row_count": N}	Transparent — bridge's reshape_duckdb_result maps back to the CLI's rows-array shape
#28	duckdb non-SELECT result envelope	Literal string {"status": "ok"}	data: {"affected_rows": N}	Transparent — bridge maps back; affected_rows dropped (CLI never exposed it)
#28	duckdb params binding	_params: &[String] accepted but silently ignored	Bound as JSON values at ? placeholders	Feature gain — playbooks that intended their params: field would now see them applied (no breakage for playbooks with stale params + no ? placeholders since DuckDB ignores extra params)
#28	duckdb path resolution + mkdir	resolve_duckdb_path + fs::create_dir_all(parent) in execute_duckdb_query	Open as-given, no mkdir	CLI keeps owning these at the call site before dispatch; bridge has no knowledge of the playbook directory

Stats since the workspace landed

• playbook_runner.rs: 2,688 → 1,606 lines (-1,082 net across PR-2a + PR-2b + PR-2c-3 + PR-2c-4 + PR-2c-5 + PR-2c-6 + PR-2c-7 + PR-2c-8)
• New code in noetl-executor: 7 modules + 1 worker submodule, ~3,400 LoC
• noetl-executor unit tests: 0 → 80 across PR-1 + PR-2a + PR-2b + PR-2c-1 + PR-2c-2 + PR-2c-3 + PR-2c-4 + PR-2c-5 + PR-2c-6 + PR-2c-7 + PR-2c-8
• Workspace-wide tests: 162 passing (80 noetl-executor + 41 noetl + 41 ntl)