Fix merge_yearly.py: stream one year at a time to avoid OOM

scripts/merge_yearly.py

@@ -5,12 +5,15 @@ merge_yearly.py
 Combines each asset's per-year Parquet files on HuggingFace into one
 big combined Parquet per asset, then uploads it back to the dataset repo.
 
+Memory-efficient: streams one year at a time using ParquetWriter,
+never holds more than one year in RAM.
+
 Assets handled: DOGEUSDT, XRPUSDT, SOLUSDT
 (BTCUSDT / ETHUSDT / BNBUSDT already have combined files)
 
-HF layout assumed:
-  data/{SYMBOL}/{SYMBOL}_{YEAR}.parquet  (source yearly files)
-  data/{SYMBOL}_1s.parquet               (output combined file)
+HF layout:
+  source:  data/{SYMBOL}/{SYMBOL}_{YEAR}.parquet
+  output:  data/{SYMBOL}_1s.parquet
 
 Usage:
   HF_TOKEN=<token> python scripts/merge_yearly.py
@@ -42,6 +45,9 @@ PA_SCHEMA = pa.schema([
     ("volume",      pa.float64()),
 ])
 
+# Write in chunks so we never have more than ~50M rows in RAM at once
+CHUNK_ROWS = 10_000_000
+
 
 def yearly_files_for(symbol: str) -> list[str]:
     """Return sorted list of HF repo paths for a symbol's yearly parquets."""
@@ -52,6 +58,7 @@ def yearly_files_for(symbol: str) -> list[str]:
 
 def merge_symbol(symbol: str, api: HfApi, dry_run: bool = False) -> None:
     hf_out_path = f"data/{symbol}_1s.parquet"
+    local_out   = Path(f"/tmp/{symbol}_1s.parquet")
     print(f"\n{'─'*55}")
     print(f"  {symbol}")
 
@@ -64,49 +71,44 @@ def merge_symbol(symbol: str, api: HfApi, dry_run: bool = False) -> None:
     for yf in yearly:
         print(f"    {yf}")
 
-    frames = []
-    for yf in yearly:
-        year = yf.split("_")[-1].replace(".parquet", "")
-        print(f"  Downloading {year}...", flush=True)
-        local = hf_hub_download(
-            repo_id=REPO_ID,
-            filename=yf,
-            repo_type="dataset",
-            token=HF_TOKEN,
-        )
-        df = pd.read_parquet(local, engine="pyarrow")
-        print(f"    {len(df):>12,} rows  ({df['open_time_s'].min()} → {df['open_time_s'].max()})")
-        frames.append(df)
-
-    print("  Concatenating...", flush=True)
-    combined = pd.concat(frames, ignore_index=True)
-    del frames; gc.collect()
-
-    before = len(combined)
-    combined = (
-        combined
-        .sort_values("open_time_s")
-        .drop_duplicates("open_time_s")
-        .reset_index(drop=True)
-    )
-    after = len(combined)
-    if before != after:
-        print(f"  Dropped {before - after:,} duplicate rows.")
-
-    print(f"  Total rows: {len(combined):,}")
-    print(f"  Date range: {combined['open_time_s'].min()} → {combined['open_time_s'].max()}")
-
     if dry_run:
-        print("  [dry-run] skipping upload.")
+        print("  [dry-run] skipping download and upload.")
         return
 
-    local_out = Path(f"/tmp/{symbol}_1s.parquet")
-    print(f"  Writing parquet...", flush=True)
-    table = pa.Table.from_pandas(combined, schema=PA_SCHEMA, preserve_index=False)
-    del combined; gc.collect()
-    pq.write_table(table, local_out, compression="snappy")
+    total_rows = 0
+    writer = None
+
+    try:
+        for yf in yearly:
+            year = yf.split("_")[-1].replace(".parquet", "")
+            print(f"  Downloading {year}...", flush=True)
+            local = hf_hub_download(
+                repo_id=REPO_ID,
+                filename=yf,
+                repo_type="dataset",
+                token=HF_TOKEN,
+            )
+
+            # Read and write in chunks to keep memory low
+            pf = pq.ParquetFile(local)
+            year_rows = 0
+            for batch in pf.iter_batches(batch_size=CHUNK_ROWS, schema=PA_SCHEMA):
+                if writer is None:
+                    writer = pq.ParquetWriter(str(local_out), PA_SCHEMA, compression="snappy")
+                writer.write_batch(batch)
+                year_rows += batch.num_rows
+
+            total_rows += year_rows
+            print(f"    {year_rows:>12,} rows written  (running total: {total_rows:,})")
+            del pf; gc.collect()
+
+    finally:
+        if writer:
+            writer.close()
+
     file_mb = local_out.stat().st_size / 1e6
-    print(f"  File size: {file_mb:.1f} MB")
+    print(f"  Total rows: {total_rows:,}")
+    print(f"  File size:  {file_mb:.1f} MB")
 
     print(f"  Uploading to {hf_out_path}...", flush=True)
     api.upload_file(
@@ -114,7 +116,7 @@ def merge_symbol(symbol: str, api: HfApi, dry_run: bool = False) -> None:
         path_in_repo=hf_out_path,
         repo_id=REPO_ID,
         repo_type="dataset",
-        commit_message=f"Merge yearly files → {symbol}_1s.parquet ({file_mb:.0f} MB)",
+        commit_message=f"Merge yearly files → {symbol}_1s.parquet ({total_rows:,} rows, {file_mb:.0f} MB)",
     )
     local_out.unlink(missing_ok=True)
     print(f"  Done — uploaded {hf_out_path}")
@@ -123,7 +125,7 @@ def merge_symbol(symbol: str, api: HfApi, dry_run: bool = False) -> None:
 def main():
     parser = argparse.ArgumentParser(description="Merge yearly Parquet files per asset")
     parser.add_argument("--dry-run", action="store_true",
-                        help="Print stats but skip uploading")
+                        help="Print stats but skip downloading and uploading")
     parser.add_argument("--symbols", nargs="+", default=SYMBOLS,
                         help="Override which symbols to process")
     args = parser.parse_args()
@@ -138,7 +140,7 @@ def main():
     print("  merge_yearly.py — Combine yearly Parquets")
     print(f"  Symbols: {', '.join(args.symbols)}")
     if args.dry_run:
-        print("  Mode: DRY RUN (no uploads)")
+        print("  Mode: DRY RUN")
     print(f"{'='*55}")
 
     for symbol in args.symbols: