From ff27871108ab6d1669c42d6fcd3949ba4dccfcf6 Mon Sep 17 00:00:00 2001
From: "google-labs-jules[bot]"
 <161369871+google-labs-jules[bot]@users.noreply.github.com>
Date: Wed, 8 Jul 2026 09:32:34 +0000
Subject: [PATCH] =?UTF-8?q?=E2=9A=A1=20Bolt:=20Replace=20pandas=20iterrows?=
 =?UTF-8?q?=20with=20to=5Fdict('records')=20for=20performance?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Co-authored-by: alinelena <3306823+alinelena@users.noreply.github.com>
---
 .jules/bolt.md                               | 4 ++++
 src/lavello_mlips/verify_processed_omol25.py | 7 ++++---
 2 files changed, 8 insertions(+), 3 deletions(-)

diff --git a/.jules/bolt.md b/.jules/bolt.md
index d87bd6d..54a2518 100644
--- a/.jules/bolt.md
+++ b/.jules/bolt.md
@@ -8,3 +8,7 @@
 ## 2024-03-29 - ASE Custom JSON encoding vs standard JSON
 **Learning:** ASE's custom JSON encoder (`ase.io.jsonio.encode`) will generate dicts with special keys like `__ndarray__` or `__complex__` (e.g. `{"__ndarray__": [[5], "int64", ...]}`). When optimizing JSON deserialization using faster alternatives like `orjson`, it's critical to realize that a normal `json.loads` or `orjson.loads` will deserialize this into a Python dictionary, while ASE's custom `decode` will properly reconstruct the underlying numpy array. Bypassing ASE's decoder without checking for these keys leads to downstream type errors (e.g. `KeyError: '__ndarray__'`).
 **Action:** When replacing or wrapping ASE's jsonio with `orjson`, always fall back to ASE's `decode` if the payload string contains `__ndarray__` or `__complex__` markers, to ensure custom objects are correctly reconstructed.
+
+## 2024-05-18 - Replacing `df.iterrows()` with `df.to_dict('records')` for large datasets
+**Learning:** In pipelines handling large dataframes, iterating through rows using `df.iterrows()` creates significant bottlenecks by instantiating new pandas `Series` objects for every row. Converting the DataFrame to a list of dicts using `df.to_dict('records')` first provides massive performance gains when row-level iteration or list/dictionary comprehension is required.
+**Action:** Always replace `df.iterrows()` with `df.to_dict('records')` in performance-critical sections iterating over pandas DataFrames, and update downstream property accesses accordingly (since the row object is now a native Python dictionary instead of a pandas Series).
diff --git a/src/lavello_mlips/verify_processed_omol25.py b/src/lavello_mlips/verify_processed_omol25.py
index 9d1c47c..d1987cd 100644
--- a/src/lavello_mlips/verify_processed_omol25.py
+++ b/src/lavello_mlips/verify_processed_omol25.py
@@ -77,8 +77,9 @@ def main() -> None:
         )
 
     logger.info(f"Loaded {len(df)} records from Parquet.")
-    parquet_by_sha = {row["geom_sha1"]: row for _, row in df.iterrows()}
-    parquet_by_argone_rel = {row["argonne_rel"]: row for _, row in df.iterrows()}
+    df_records = df.to_dict("records")
+    parquet_by_sha = {row["geom_sha1"]: row for row in df_records}
+    parquet_by_argone_rel = {row["argonne_rel"]: row for row in df_records}
     logger.info(f"Loading ExtXYZ file from {args.extxyz} (this may take a moment)...")
     all_atoms = read(str(args.extxyz), index=":")
     if not isinstance(all_atoms, list):
@@ -108,7 +109,7 @@ def get_dump_entry(at):
             info = dict(at.info)
             rel = info.get("argonne_rel")
             pq_row = parquet_by_argone_rel.get(rel)
-            pq_data = pq_row.to_dict() if pq_row is not None else None
+            pq_data = pq_row if pq_row is not None else None
             return {"xyz": info, "parquet": pq_data}
 
         duplicates = {