diff --git a/src/ptscrapper/scheduler.py b/src/ptscrapper/scheduler.py
index bdc8f91..c28a2d2 100644
--- a/src/ptscrapper/scheduler.py
+++ b/src/ptscrapper/scheduler.py
@@ -32,7 +32,7 @@ from .storage import (
     push_to_nas,
     write_course_geometry,
     write_course_posts,
-    write_positions,
+    write_positions_batched,
     write_posts_lines,
     write_stops,
 )
@@ -134,7 +134,7 @@ async def positions_job():
         table = pa.Table.from_pylist(enriched)
 
         # 4) pass both timestamps through
-        write_positions(table, server_ts)
+        write_positions_batched(table, server_ts)
         logger.info("wrote positions shard for %s", server_ts)
     except Exception:
         # record error details
diff --git a/src/ptscrapper/storage.py b/src/ptscrapper/storage.py
index aa22a25..710abcb 100644
--- a/src/ptscrapper/storage.py
+++ b/src/ptscrapper/storage.py
@@ -2,6 +2,7 @@
 
 import glob
 import os
+import threading
 from datetime import date, datetime
 
 import geoarrow.pyarrow as ga
@@ -18,29 +19,64 @@ from .config import (  # base path on the remote, e.g. "pt-scraper-data"
     RCLONE_REMOTE_PATH,
 )
 
+# module‐level state for our rolling 10-min writer
+_positions_writer = None  # pq.ParquetWriter
+_positions_batch = None  # datetime for current batch start
+_positions_path = None  # where we’re writing right now
+_positions_lock = threading.Lock()
+
 
 def _ensure_dir(path: str):
     """Make sure the directory exists."""
     os.makedirs(path, exist_ok=True)
 
 
-def write_positions(
-    table: pa.Table,
-    server_ts: datetime,
-) -> str:
+def write_positions_batched(table: pa.Table, server_ts: datetime) -> None:
     """
-    Write out the raw positions (including x,y, server_ts, fetched_ts).
-    Geometry will be added later during daily compaction.
+    Append `table` into a rolling Parquet file, one file per 10-minute window.
+    When the window rolls over, we close+push the old file and start a new one.
     """
+    global _positions_writer, _positions_batch, _positions_path
 
-    subdir = server_ts.strftime("positions/%Y/%m/%d/%H")
-    local_dir = os.path.join(LOCAL_DATA_DIR, subdir)
-    _ensure_dir(local_dir)
+    # compute the ten-minute bucket start (floor to minute//10 * 10)
+    batch_start = server_ts.replace(
+        minute=(server_ts.minute // 10) * 10, second=0, microsecond=0
+    )
 
-    filename = f"{server_ts.strftime('%M%S')}_{HOST_ID}.parquet"
-    local_path = os.path.join(local_dir, filename)
-    pq.write_table(table, local_path, compression="snappy")
-    return local_path
+    with _positions_lock:
+        # if we’ve moved into a new bucket, close & push the old
+        if _positions_batch is None or batch_start != _positions_batch:
+            if _positions_writer is not None:
+                _positions_writer.close()
+                # push the completed file
+                rel = os.path.relpath(_positions_path, LOCAL_DATA_DIR).replace(
+                    os.sep, "/"
+                )
+                rclone.copy(
+                    _positions_path,
+                    f"{RCLONE_REMOTE}:{RCLONE_REMOTE_PATH}/{rel}",
+                    ignore_existing=True,
+                )
+
+            # start a new writer for the new bucket
+            subdir = batch_start.strftime("positions/%Y/%m/%d/%H")
+            local_dir = os.path.join(LOCAL_DATA_DIR, subdir)
+            _ensure_dir(local_dir)
+            fn = f"{batch_start.strftime('%M%S')}_{HOST_ID}.parquet"
+            _positions_path = os.path.join(local_dir, fn)
+            _positions_writer = pq.ParquetWriter(
+                _positions_path,
+                table.schema,
+                compression="snappy",
+                use_dictionary=False,
+                write_statistics=False,
+            )
+            _positions_batch = batch_start
+
+        # write this 10-s table into the current bucket’s file
+        _positions_writer.write_table(table)
+
+    # no return path: we push only on rollover
 
 
 def write_course_posts(table: pa.Table, course_id: int, ts: datetime) -> str: