From 1fb4f5857590605d581faf1033b6ee98218a9673 Mon Sep 17 00:00:00 2001
From: Simone <simone.chifari@gmail.com>
Date: Wed, 10 Jun 2026 00:18:25 +0200
Subject: [PATCH] perf(winrm): free streamed PSRP records to cap host RAM
 during prepare
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

run_streaming used index-based _drain, which read new records from
ps.output / ps.streams.{information,error} but never released the consumed
ones. pypsrp retains every record (each a heavy deserialised PSObject) for
the whole invocation, so a long prepare-win (Windows Update + toolchain,
thousands of lines) grew the host python process RAM without bound.

Replace _drain with _drain_clear, which emits each pending record then
deletes the consumed prefix (`del seq[:n]`). Safe because draining runs on
the same thread as poll_invoke — nothing appends while we mutate. Track an
explicit had_error flag since the error stream is now cleared as we go.
---
 src/ci_orchestrator/transport/winrm.py | 43 +++++++++++++++++---------
 1 file changed, 29 insertions(+), 14 deletions(-)

diff --git a/src/ci_orchestrator/transport/winrm.py b/src/ci_orchestrator/transport/winrm.py
index 5ef4ff0..d9fa6f1 100644
--- a/src/ci_orchestrator/transport/winrm.py
+++ b/src/ci_orchestrator/transport/winrm.py
@@ -51,11 +51,23 @@ def _info_text(record: Any) -> str:
     return str(data) if data is not None else str(record)
 
 
-def _drain(seq: list[Any], start: int, emit: Callable[[str], None]) -> int:
-    """Emit ``str()`` of every item in ``seq[start:]``; return the new length."""
-    for item in seq[start:]:
-        emit(str(item))
-    return len(seq)
+def _drain_clear(
+    seq: list[Any], render: Callable[[Any], str], emit: Callable[[str], None]
+) -> None:
+    """Emit ``render(item)`` for each item in ``seq`` then drop those items.
+
+    Clearing the consumed prefix is essential: pypsrp keeps every output/info/
+    error record in these lists for the whole invocation, and each is a heavy
+    deserialised PSObject. A long prepare (Windows Update + toolchain, thousands
+    of lines) would otherwise grow the host process RAM without bound. We run on
+    the same thread as ``poll_invoke`` so nothing appends while we mutate.
+    """
+    if not seq:
+        return
+    n = len(seq)
+    for item in seq[:n]:
+        emit(render(item))
+    del seq[:n]
 
 
 class WinRmTransport:
@@ -224,20 +236,23 @@ class WinRmTransport:
             ps = PowerShell(pool)
             ps.add_script(script)
             ps.begin_invoke()
-            out_i = info_i = err_i = 0
             running = (PSInvocationState.NOT_STARTED, PSInvocationState.RUNNING)
+            had_error = False
             while ps.state in running:
                 ps.poll_invoke()
-                out_i = _drain(ps.output, out_i, emit_out)
-                info_i = _drain(
-                    [_info_text(r) for r in ps.streams.information], info_i, on_line
-                )
-                err_i = _drain([str(e) for e in ps.streams.error], err_i, emit_err)
+                _drain_clear(ps.output, str, emit_out)
+                _drain_clear(ps.streams.information, _info_text, on_line)
+                if ps.streams.error:
+                    had_error = True
+                _drain_clear(ps.streams.error, str, emit_err)
                 time.sleep(0.5)
             ps.end_invoke()
-            _drain(ps.output, out_i, emit_out)
-            _drain([str(e) for e in ps.streams.error], err_i, emit_err)
-            returncode = 1 if ps.streams.error else 0
+            _drain_clear(ps.output, str, emit_out)
+            _drain_clear(ps.streams.information, _info_text, on_line)
+            if ps.streams.error:
+                had_error = True
+            _drain_clear(ps.streams.error, str, emit_err)
+            returncode = 1 if had_error else 0
         finally:
             with contextlib.suppress(Exception):
                 pool.close()