bundle: record and read deployment state via DMS

acceptance/bin/print_requests.py

@@ -178,6 +178,17 @@ def filter_requests(requests, path_filters, include_get, should_sort, unique=Fal
     return filtered_requests
 
 
+def del_path(body, field):
+    """Delete field from body. A dotted field descends into nested objects, e.g.
+    deployment.version_id removes only that key from the deployment block."""
+    *parents, leaf = field.split(".")
+    for name in parents:
+        body = body.get(name)
+        if not isinstance(body, dict):
+            return
+    body.pop(leaf, None)
+
+
 def main():
     parser = argparse.ArgumentParser()
     parser.add_argument("path_filters", nargs="*", help="Path substring filters")
@@ -238,7 +249,7 @@ def main():
         body = req.get("body")
         if isinstance(body, dict):
             for field in del_body_fields:
-                body.pop(field, None)
+                del_path(body, field)
         for field in del_fields:
             req.pop(field, None)
     if args.verbose:

acceptance/bundle/ai_runtime_task/test.toml

@@ -0,0 +1,55 @@
+# These normalize the deployment stamp recording adds to every job and pipeline, so a test
+# asserts the same golden files whether or not recording is on - that is the point of the
+# DMS run, rather than keeping a second copy of 600-odd output files. They are repeated per
+# subtree rather than living in the parent because bundle/dms asserts the stamp itself and
+# would inherit them.
+#
+# The stamp as a change the plan reports on its own. It shows up at whatever depth the
+# enclosing object sits at, so the indent is matched loosely; the body lines are matched as
+# `"key": value` pairs rather than `.*` so the match stops at the entry's own closing brace
+# instead of running into its siblings. (Go's regexp is RE2, so the indent cannot be
+# captured and back-referenced.)
+[[Repls]]
+Old = '(?m)^ *"deployment\.(deployment_id|version_id)": \{\n(?: *"[a-z_]+": .*\n)+ *\},\n'
+New = ''
+
+# Same entry when it is the last one in the object, so the comma is on the line before.
+[[Repls]]
+Old = '(?m),\n *"deployment\.(deployment_id|version_id)": \{\n(?: *"[a-z_]+": .*\n)+ *\}\n'
+New = "\n"
+
+# When the stamp is the only entry, the whole "changes" object exists because of recording.
+# The trailing-comma form comes first: the rule after it would match the same text and
+# leave the comma on the previous line dangling.
+[[Repls]]
+Old = '(?m),\n *"changes": \{\n *"deployment\.(deployment_id|version_id)": \{\n(?: *"[a-z_]+": .*\n)+ *\}\n *\}\n'
+New = "\n"
+
+[[Repls]]
+Old = '(?m)^ *"changes": \{\n *"deployment\.(deployment_id|version_id)": \{\n(?: *"[a-z_]+": .*\n)+ *\}\n *\},?\n'
+New = ''
+
+# The stamp inside a serialized deployment block, as printed by `jobs get` / `pipelines get`
+# and in state dumps. Both keys always sit alongside "kind" and "metadata_file_path", so
+# each pattern anchors on one of those - that keeps it from matching an unrelated field
+# named deployment_id elsewhere in the output.
+#
+# Order puts these after the root's numeric rules (Order = 10), which have by then turned
+# the id into [NUMID].
+[[Repls]]
+Old = '(?m)^ *"deployment_id": "\[NUMID\]",\n( *"kind": "BUNDLE")'
+New = '${1}'
+Order = 20
+
+[[Repls]]
+# The value is required to be non-empty: a terraform state dump carries
+# "version_id": "" for a job it never stamped, and that line is not ours to drop.
+Old = '(?m)(^ *"metadata_file_path": [^\n]*),\n *"version_id": "[^"\n]+"\n'
+New = "$1\n"
+Order = 20
+
+# Same two keys in gron.py's flattened form, where each is its own line.
+[[Repls]]
+Old = '(?m)^json.*\.deployment\.(deployment_id|version_id) = [^\n]*\n'
+New = ''
+Order = 20

acceptance/bundle/artifacts/test.toml

@@ -29,3 +29,59 @@ Response.Body = '''
   "spark_version": "13.3.x-scala2.12"
 }
 '''
+
+# These normalize the deployment stamp recording adds to every job and pipeline, so a test
+# asserts the same golden files whether or not recording is on - that is the point of the
+# DMS run, rather than keeping a second copy of 600-odd output files. They are repeated per
+# subtree rather than living in the parent because bundle/dms asserts the stamp itself and
+# would inherit them.
+#
+# The stamp as a change the plan reports on its own. It shows up at whatever depth the
+# enclosing object sits at, so the indent is matched loosely; the body lines are matched as
+# `"key": value` pairs rather than `.*` so the match stops at the entry's own closing brace
+# instead of running into its siblings. (Go's regexp is RE2, so the indent cannot be
+# captured and back-referenced.)
+[[Repls]]
+Old = '(?m)^ *"deployment\.(deployment_id|version_id)": \{\n(?: *"[a-z_]+": .*\n)+ *\},\n'
+New = ''
+
+# Same entry when it is the last one in the object, so the comma is on the line before.
+[[Repls]]
+Old = '(?m),\n *"deployment\.(deployment_id|version_id)": \{\n(?: *"[a-z_]+": .*\n)+ *\}\n'
+New = "\n"
+
+# When the stamp is the only entry, the whole "changes" object exists because of recording.
+# The trailing-comma form comes first: the rule after it would match the same text and
+# leave the comma on the previous line dangling.
+[[Repls]]
+Old = '(?m),\n *"changes": \{\n *"deployment\.(deployment_id|version_id)": \{\n(?: *"[a-z_]+": .*\n)+ *\}\n *\}\n'
+New = "\n"
+
+[[Repls]]
+Old = '(?m)^ *"changes": \{\n *"deployment\.(deployment_id|version_id)": \{\n(?: *"[a-z_]+": .*\n)+ *\}\n *\},?\n'
+New = ''
+
+# The stamp inside a serialized deployment block, as printed by `jobs get` / `pipelines get`
+# and in state dumps. Both keys always sit alongside "kind" and "metadata_file_path", so
+# each pattern anchors on one of those - that keeps it from matching an unrelated field
+# named deployment_id elsewhere in the output.
+#
+# Order puts these after the root's numeric rules (Order = 10), which have by then turned
+# the id into [NUMID].
+[[Repls]]
+Old = '(?m)^ *"deployment_id": "\[NUMID\]",\n( *"kind": "BUNDLE")'
+New = '${1}'
+Order = 20
+
+[[Repls]]
+# The value is required to be non-empty: a terraform state dump carries
+# "version_id": "" for a job it never stamped, and that line is not ours to drop.
+Old = '(?m)(^ *"metadata_file_path": [^\n]*),\n *"version_id": "[^"\n]+"\n'
+New = "$1\n"
+Order = 20
+
+# Same two keys in gron.py's flattened form, where each is its own line.
+[[Repls]]
+Old = '(?m)^json.*\.deployment\.(deployment_id|version_id) = [^\n]*\n'
+New = ''
+Order = 20