\n| \u9002\u7528<\/td>\n | \u63a2\u7d22\u6027\u5206\u6790\u3001\u673a\u5668\u5b66\u4e60\u7279\u5f81\u5de5\u7a0b\u3001\u751f\u6001\u4f9d\u8d56<\/td>\n | \u5927\u6279\u91cf\u6570\u636e\u6e05\u6d17\u3001ETL\u3001\u62a5\u8868\u805a\u5408\u3001\u6d41\u5f0f\u5904\u7406<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
\n2. \u6838\u5fc3\u6982\u5ff5\u901f\u89c8<\/h3>\n\u5148\u5efa\u7acb Polars \u7684\u4e16\u754c\u89c2，\u5b83\u53ea\u6709\u51e0\u4e2a\u5173\u952e\u6982\u5ff5，\u4f46\u6bcf\u4e2a\u90fd\u6bd4 Pandas \u66f4\u8f83\u771f\u3002<\/p>\n 2.1 DataFrame \u4e0e Series（\u548c Pandas \u540c\u540d\u4f46\u4e0d\u540c\u5185\u6838）<\/h4>\n\n- pl.DataFrame：\u4e8c\u7ef4\u8868，\u5e95\u5c42\u662f\u00a0Arrow \u5217\u5f0f\u5b58\u50a8，\u6bcf\u4e00\u5217\u662f\u4e00\u4e2a\u8fde\u7eed\u5185\u5b58\u6570\u7ec4，\u7c7b\u578b\u4e00\u81f4；<\/li>\n
- pl.Series：\u5355\u5217，\u672c\u8d28\u662f\u6307\u5411 Arrow ChunkedArray \u7684\u53e5\u67c4；<\/li>\n
- \u5217\u7684\u7c7b\u578b\u6709\u4e25\u683c\u8bed\u4e49：Int8\/16\/32\/64\u3001UInt*\u3001Float32\/64\u3001String\u3001Boolean\u3001Date\/Datetime\/Duration\u3001List\u3001Struct\u3001Categorical\u3001Enum\u3002<\/li>\n<\/ul>\n
2.2 \u8868\u8fbe\u5f0f（Expression）：Polars \u7684\u7075\u9b42<\/h4>\nPandas \u7684 API \u662f\u65b9\u6cd5\u4f5c\u7528\u5728 DataFrame \u4e0a；Polars \u7684 API \u662f\u00a0\u8868\u8fbe\u5f0f\u4f5c\u7528\u5728\u5217\u4e0a，\u518d\u4f20\u7ed9\u4e0a\u4e0b\u6587\u3002<\/p>\n # Pandas \u98ce\u683c：\u65b9\u6cd5\u5206\u6563
\ndf[df['ts'].dt.hour == 8]['value'].mean()<\/p>\n # Polars \u98ce\u683c：\u8868\u8fbe\u5f0f\u7edf\u4e00
\ndf.filter(pl.col('ts').dt.hour() == 8).select(pl.col('value').mean())<\/p>\n \u8868\u8fbe\u5f0f\u53ef\u4ee5\u7ec4\u5408\u3001\u590d\u7528\u3001\u5e76\u884c\u6c42\u503c\u3002\u4e00\u4e2a\u8868\u8fbe\u5f0f\u5bf9\u8c61\u53ef\u4ee5\u5728\u591a\u4e2a\u4e0a\u4e0b\u6587\u91cc\u7528：select（\u9009\u5217）\u3001filter（\u8fc7\u6ee4）\u3001with_columns（\u65b0\u589e\u5217）\u3001group_by（\u5206\u7ec4\u805a\u5408）\u3001sort（\u6392\u5e8f）\u3002<\/p>\n 2.3 Eager \u4e0e Lazy \u53cc\u6a21\u5f0f<\/h4>\n\n- Eager：.collect()\u00a0\u524d\u7684 DataFrame \u64cd\u4f5c，\u6bcf\u6b65\u7acb\u523b\u6267\u884c；<\/li>\n
- Lazy：.lazy()\u00a0\u751f\u6210 LazyFrame，\u53ea\u8bb0\u5f55\u67e5\u8be2\u8ba1\u5212，\u6700\u7ec8\u00a0.collect()\u00a0\u624d\u6267\u884c\u3002\u4e2d\u95f4\u4e0d\u4f1a\u4ea7\u751f\u4efb\u4f55\u4e34\u65f6\u5e27，\u4e14\u4f1a\u89e6\u53d1\u67e5\u8be2\u4f18\u5316\u5668\u3002<\/li>\n<\/ul>\n
\u8fd9\u662f Polars \u6027\u80fd\u795e\u8bdd\u7684\u6838\u5fc3\u4e4b\u4e00：lazy\u00a0\u6a21\u5f0f\u4f1a\u628a\u4f60\u7684\u6574\u6761\u6d41\u6c34\u7ebf（\u8bfb\u6587\u4ef6 \u2192 \u8fc7\u6ee4 \u2192 \u805a\u5408 \u2192 \u6392\u5e8f）\u5408\u5e76\u4f18\u5316\u540e\u518d\u6267\u884c，\u800c\u4e0d\u662f\u50cf Pandas \u4e00\u6837\u6bcf\u4e00\u6b65\u5efa\u4e00\u4e2a\u4e34\u65f6\u8868\u3002<\/p>\n 2.4 \u5217\u5f0f\u5b58\u50a8\u4e3a\u4ec0\u4e48\u5feb<\/h4>\n\u6570\u636e\u5728\u5185\u5b58\u4e2d\u662f\u6309\u5217\u8fde\u7eed\u5b58\u653e\u7684\u3002\u8fd9\u610f\u5473\u7740：<\/p>\n \n- \u8bfb\u53d6\u4e00\u5217 = \u987a\u5e8f\u626b\u63cf\u4e00\u6bb5\u8fde\u7eed\u5185\u5b58，CPU \u7f13\u5b58\u547d\u4e2d\u7387\u9ad8；<\/li>\n
- \u8fc7\u6ee4\u4e00\u5217\u65f6，\u5176\u4ed6\u5217\u5b8c\u5168\u4e0d\u78b0（Pandas \u884c\u5f0f\u5b58\u50a8\u8981\u6574\u884c\u8bfb\u51fa\u518d\u4e22\u5f03）；<\/li>\n
- \u6570\u503c\u8ba1\u7b97\u5929\u7136\u5411\u91cf\u5316（\u4e00\u5217\u6570\u636e\u5c31\u662f\u4e00\u7247\u540c\u7c7b\u578b\u6570\u7ec4，\u53ef\u4ee5\u76f4\u63a5\u4e0a SIMD）；<\/li>\n
- \u538b\u7f29\u53cb\u597d（\u540c\u7c7b\u578b\u6570\u636e\u538b\u7f29\u7387\u9ad8，pl.read_parquet\u00a0\u662f\u4e3b\u6d41\u683c\u5f0f）\u3002<\/li>\n<\/ul>\n
2.5 \u6d41\u5f0f\u5904\u7406（Streaming）<\/h4>\n\u9762\u5bf9\u300c\u8d85\u8fc7\u5185\u5b58\u300d\u7684\u6570\u636e，Polars \u63d0\u4f9b\u6d41\u5f0f\u6267\u884c：.collect(streaming=True)，\u67e5\u8be2\u5f15\u64ce\u4f1a\u5728\u5fc5\u8981\u65f6\u628a\u4e2d\u95f4\u7ed3\u679c\u843d\u76d8（\u5982\u6392\u5e8f\u3001\u805a\u5408），\u7528\u8fb9\u8bfb\u8fb9\u7b97\u4ee3\u66ff\u5168\u91cf\u8f7d\u5165\u5185\u5b58\u3002<\/p>\n
\n3. \u5b89\u88c5\u4e0e\u5feb\u901f\u4e0a\u624b<\/h3>\npip install polars # \u5f53\u524d\u7a33\u5b9a\u7248 1.x
\npip install pyarrow # \u53ef\u9009\u7684 Arrow \u4e92\u64cd\u4f5c\u6269\u5c55
\npip install polars[timezone] # \u65f6\u533a\u652f\u6301<\/p>\n Polars \u662f\u7f16\u8bd1\u597d\u7684\u4e8c\u8fdb\u5236 wheel，\u4e0d\u9700\u8981\u672c\u5730 Rust \u73af\u5883\u3002\u5bfc\u5165：<\/p>\n import polars as pl<\/p>\n \u4e94\u5206\u949f\u6700\u5c0f\u95ed\u73af：<\/p>\n # \u6784\u9020
\ndf = pl.DataFrame({
\n "device": ["CNC-01", "CNC-01", "CNC-02", "CNC-02"],
\n "ts": ["2026-10-23 08:00:00", "2026-10-23 08:00:01",
\n "2026-10-23 08:00:00", "2026-10-23 08:00:01"],
\n "value": [10.5, 12.1, 8.2, 9.0],
\n}).with_columns(pl.col("ts").str.to_datetime())<\/p>\n # \u8fc7\u6ee4 + \u805a\u5408（Polars \u98ce\u683c）
\nresult = (
\n df.lazy()
\n .filter(pl.col("value") > 9)
\n .group_by("device")
\n .agg(pl.col("value").mean().alias("avg_value"))
\n .sort("avg_value", descending=True)
\n .collect()
\n)
\nprint(result)<\/p>\n \u8f93\u51fa：<\/p>\n shape: (2, 2)
\n\u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510
\n\u2502 device \u2506 avg_value \u2502
\n\u255e\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u256a\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2561
\n\u2502 CNC-01 \u2506 11.3 \u2502
\n\u2502 CNC-02 \u2506 8.2 \u2502
\n\u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518<\/p>\n
\n4. API \u8bf4\u660e（\u9ad8\u9891\u901f\u67e5）<\/h3>\n4.1 \u521b\u5efa\u4e0e IO<\/h4>\n\n\n\u76ee\u7684API<\/tr>\n \n| \u4ece dict \/ list \/ numpy \u6784\u9020<\/td>\n | pl.DataFrame({…})\u3001pl.DataFrame(arr, schema=…)<\/td>\n<\/tr>\n | \n| \u8bfb CSV<\/td>\n | pl.read_csv(path, schema_overrides={…}, try_parse_dates=True)<\/td>\n<\/tr>\n | \n| \u8bfb Parquet \/ Arrow IPC<\/td>\n | pl.read_parquet(path)\u3001pl.read_ipc(path)<\/td>\n<\/tr>\n | \n| \u8bfb JSON \/ NDJSON<\/td>\n | pl.read_json(path)\u3001pl.read_ndjson(path)<\/td>\n<\/tr>\n | \n| \u8fde\u63a5\u6570\u636e\u5e93<\/td>\n | pl.read_database(uri=…, query=…)<\/td>\n<\/tr>\n | \n| \u5199\u6587\u4ef6<\/td>\n | df.write_csv(…)\u3001df.write_parquet(…)\u3001df.write_json(…)\u3001df.write_ipc(…)<\/td>\n<\/tr>\n | \n| \u4e0e Pandas \/ NumPy \u4e92\u8f6c<\/td>\n | df.to_pandas()\u3001df.to_numpy()\u3001pl.from_pandas(pd_df)\u3001pl.from_numpy(arr)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n4.2 \u9009\u62e9\u4e0e\u64cd\u4f5c\u4e0a\u4e0b\u6587<\/h4>\n\n\n\u4e0a\u4e0b\u6587\u4f5c\u7528\u793a\u4f8b<\/tr>\n \n| df.select(…)<\/td>\n | \u9009\u5217 \/ \u8ba1\u7b97\u5217<\/td>\n | df.select(pl.col("a"), (pl.col("b")*2).alias("b2"))<\/td>\n<\/tr>\n | \n| df.with_columns(…)<\/td>\n | \u65b0\u589e \/ \u8986\u76d6\u5217<\/td>\n | df.with_columns(pl.col("v").diff().alias("delta"))<\/td>\n<\/tr>\n | \n| df.filter(…)<\/td>\n | \u884c\u8fc7\u6ee4<\/td>\n | df.filter(pl.col("q") == 1)<\/td>\n<\/tr>\n | \n| df.group_by(…).agg(…)<\/td>\n | \u5206\u7ec4\u805a\u5408<\/td>\n | df.group_by("device").agg(pl.len(), pl.col("v").mean())<\/td>\n<\/tr>\n | \n| df.sort(…)<\/td>\n | \u6392\u5e8f<\/td>\n | df.sort("ts")\u3001df.sort(pl.col("v").rank())<\/td>\n<\/tr>\n | \n| df.unique()<\/td>\n | \u53bb\u91cd<\/td>\n | df.unique(subset=["device"])<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n4.3 \u5217\u9009\u62e9\u8bed\u6cd5<\/h4>\n\n- pl.col("a")\u3001pl.col("a", "b")\u3001pl.col("^prefix_.*$")（\u6b63\u5219）\u3001pl.all()\u3001pl.exclude("ts")；<\/li>\n
- df.select(pl.col("a").name.prefix("raw_"))\u00a0\u6279\u91cf\u6539\u540d；<\/li>\n
- \u53d6\u5b50\u96c6：df.select(["device", "value"])\u00a0\u76f4\u63a5\u4f20\u5b57\u7b26\u4e32\u5217\u8868\u4e5f\u53ef\u4ee5\u3002<\/li>\n<\/ul>\n
4.4 \u5e38\u7528\u8868\u8fbe\u5f0f\u51fd\u6570\u65cf<\/h4>\n\n\n\u7c7b\u522b\u8868\u8fbe\u5f0f<\/tr>\n \n| \u6570\u503c<\/td>\n | sum\/mean\/median\/std\/var\/min\/max\u3001diff\u3001pct_change\u3001rank\u3001cum_sum\u3001clip\u3001round\u3001abs\u3001log\u3001pow\u3001shift<\/td>\n<\/tr>\n | \n| \u5b57\u7b26\u4e32<\/td>\n | str.len_chars\u3001str.contains\u3001str.starts_with\u3001str.to_uppercase\u3001str.split\u3001str.extract\u3001str.replace<\/td>\n<\/tr>\n | \n| \u65f6\u95f4<\/td>\n | dt.year\/month\/day\/hour\/minute\/second\u3001dt.truncate("1h")\u3001dt.round("5m")\u3001dt.epoch("s")\u3001dt.strftime<\/td>\n<\/tr>\n | \n| \u6761\u4ef6<\/td>\n | pl.when(cond).then(a).otherwise(b)（\u5bf9\u6807\u00a0np.where）\u3001is_in\u3001is_null\u3001is_between<\/td>\n<\/tr>\n | \n| \u805a\u5408<\/td>\n | first\/last\/n_unique\/unique\/count\/len\u3001quantile(0.95)\u3001arg_max\/arg_min\u3001n_unique<\/td>\n<\/tr>\n | \n| \u5217\u8868\/\u7ed3\u6784<\/td>\n | list.len\u3001list.first\u3001list.sum\u3001list.eval\u3001struct.field<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n4.5 \u7a97\u53e3\u51fd\u6570<\/h4>\nover()\u00a0\u662f Polars \u7684\u5206\u7ec4\u7a97\u53e3\u8bed\u6cd5，\u5bf9\u6807 SQL \u7684\u00a0PARTITION BY：<\/p>\n df.with_columns(
\n pl.col("value").mean().over("device").alias("device_avg"), # \u7ec4\u5185\u5747\u503c\u5e7f\u64ad
\n pl.col("value").rank().over(["device", pl.col("ts").dt.hour()]).alias("rank_in_hour"),
\n)<\/p>\n 4.6 Join<\/h4>\ndf1.join(df2, on="device", how="left") # left\/right\/inner\/outer\/anti\/semi\/cross
\ndf1.join_asof(df2, on="ts", by="device") # \u6700\u8fd1\u65f6\u95f4\u5339\u914d，\u5de5\u4e1a\u65f6\u5e8f\u5bf9\u9f50\u795e\u5668<\/p>\n join_asof\u00a0\u662f Polars（\u53ca DuckDB）\u72ec\u6709\u4eae\u70b9：\u4e24\u8868\u6309\u65f6\u95f4\u5c31\u8fd1\u5bf9\u9f50，\u505a\u8bbe\u5907\u6863\u6848\u4e0e\u70b9\u4f4d\u660e\u7ec6\u5173\u8054\u65f6\u4e0d\u9700\u8981\u5148\u505a\u65f6\u95f4\u5206\u7bb1\u3002<\/p>\n 4.7 \u5176\u4ed6\u9ad8\u9891\u65b9\u6cd5<\/h4>\n\n- df.describe()：\u7edf\u8ba1\u6458\u8981；<\/li>\n
- df.head()\/tail()\u3001df.sample(n=1000)：\u62bd\u6837；<\/li>\n
- df.explode("list_col")：\u628a\u5217\u8868\u5217\u70b8\u5f00\u6210\u591a\u884c；<\/li>\n
- df.pivot(values="v", index="device", columns="hour")：\u900f\u89c6；<\/li>\n
- df.estimated_size()：\u4f30\u7b97\u5185\u5b58；<\/li>\n
- df.glimpse()：\u5217\u7c7b\u578b\u5feb\u901f\u9884\u89c8（\u7c7b\u4f3c\u00a0df.info()）\u3002<\/li>\n<\/ul>\n
\n5. \u8be6\u7ec6\u4f7f\u7528\u8bf4\u660e（8 \u7ec4\u53ef\u8fd0\u884c\u793a\u4f8b）<\/h3>\n\u4ee5\u4e0b\u793a\u4f8b\u56f4\u7ed5\u5de5\u4e1a\u6570\u91c7\u573a\u666f\u7ec4\u7ec7，\u5168\u90e8\u53ef\u76f4\u63a5\u8fd0\u884c\u3002<\/p>\n 5.1 \u6700\u5c0f\u95ed\u73af：\u4ece CSV \u8bfb\u70b9\u4f4d\u660e\u7ec6\u5e76\u505a\u5206\u949f\u7ea7\u805a\u5408<\/h4>\nimport polars as pl<\/p>\n # \u6a21\u62df 1 \u53f0\u8bbe\u5907 3 \u5929\u7684\u79d2\u7ea7\u70b9\u4f4d（\u8fd9\u91cc\u76f4\u63a5\u9020\u6570\u636e，\u5b9e\u9645\u7528 pl.read_csv）
\ndf = pl.DataFrame({
\n "device": ["CNC-01"] * 259200,
\n "ts": pl.datetime_range(
\n start="2026-10-20 00:00:00", end="2026-10-22 23:59:59",
\n interval="1s", eager=True,
\n ),
\n "load": [60 + (i % 40) for i in range(259200)],
\n}).with_columns(pl.col("ts").dt.cast_time_unit("ms").alias("ts_ms"))<\/p>\n minute_stats = (
\n df.lazy()
\n .with_columns(pl.col("ts").dt.truncate("1m").alias("minute"))
\n .group_by(["device", "minute"])
\n .agg(
\n pl.col("load").mean().alias("avg_load"),
\n pl.col("load").max().alias("max_load"),
\n pl.col("load").quantile(0.95).alias("p95_load"),
\n )
\n .collect()
\n)
\nprint(minute_stats.head())<\/p>\n \u8981\u70b9：dt.truncate("1m")\u00a0\u628a\u65f6\u95f4\u5bf9\u9f50\u5230\u5206\u949f；group_by\u00a0\u5185\u4e00\u6b21\u6027\u7b97\u4e09\u4e2a\u6307\u6807，\u6bd4 Pandas \u4e09\u6b21 agg \u518d concat \u7701\u5f97\u591a\u3002<\/p>\n 5.2 \u60f0\u6027\u67e5\u8be2\u4f18\u5316：\u4e09\u6b65\u6d41\u6c34\u7ebf\u5408\u5e76\u4e3a\u4e00\u6b21\u6267\u884c<\/h4>\n# \u5148 lzma，\u518d collect：\u4e2d\u95f4\u4e0d\u4ea7\u751f\u4e34\u65f6 DataFrame
\nplan = (
\n df.lazy()
\n .filter(pl.col("load") > 50)
\n .with_columns((pl.col("load") – 60).alias("overload"))
\n .group_by("device")
\n .agg(pl.col("overload").mean())
\n)
\nprint(plan.explain()) # \u67e5\u770b\u4f18\u5316\u540e\u7684\u6267\u884c\u8ba1\u5212
\nresult = plan.collect() # \u771f\u6b63\u6267\u884c<\/p>\n explain()\u00a0\u80fd\u76f4\u63a5\u770b\u5230：Polars \u4f1a\u628a\u00a0filter\u00a0\u4e0b\u63a8\u5230\u8bfb\u53d6\u9636\u6bb5（ProjectionPushdown），\u6574\u6761\u94fe\u8def\u53ea\u626b\u63cf\u4e00\u6b21\u6570\u636e\u3002<\/p>\n 5.3 \u7a97\u53e3\u51fd\u6570：\u8bbe\u5907\u65e5\u5747\u503c + \u76f8\u5bf9\u504f\u5dee<\/h4>\ndf.with_columns(
\n pl.col("ts").dt.date().alias("day"),
\n).with_columns(
\n (pl.col("load") \/ pl.col("load").mean().over(["device", "day"]) – 1).alias("deviation"),
\n).select("device", "day", "deviation").head()<\/p>\n 5.4 \u591a\u8868\u5173\u8054：\u70b9\u4f4d\u660e\u7ec6 \u00d7 \u8bbe\u5907\u6863\u6848 \u00d7 \u5de5\u827a\u53c2\u6570<\/h4>\ndetails = pl.DataFrame({…}) # \u79d2\u7ea7\u70b9\u4f4d
\ndevices = pl.DataFrame({ # \u8bbe\u5907\u6863\u6848
\n "device": ["CNC-01", "CNC-02"],
\n "line": ["L1", "L2"],
\n "model": ["VMC-850", "VMC-1060"],
\n})
\nlimits = pl.DataFrame({ # \u5de5\u827a\u53c2\u6570\u9608\u503c
\n "model": ["VMC-850", "VMC-1060"],
\n "load_hi": [85, 88],
\n})<\/p>\n enriched = (
\n details
\n .join(devices, on="device", how="left")
\n .join(limits, on="model", how="left")
\n .with_columns(
\n pl.when(pl.col("load") > pl.col("load_hi"))
\n .then(pl.lit("ALARM")).otherwise(pl.lit("OK")).alias("status")
\n )
\n)<\/p>\n 5.5 join_asof：\u6863\u6848\u53d8\u66f4\u4e0e\u70b9\u4f4d\u65f6\u5e8f\u5bf9\u9f50<\/h4>\nevents = pl.DataFrame({ # \u6362\u5200\u3001\u6539\u53c2\u7b49\u4e8b\u4ef6
\n "ts": ["2026-10-22 09:10:00", "2026-10-22 14:30:00"],
\n "device": ["CNC-01", "CNC-01"],
\n "event": ["tool_change", "param_update"],
\n}).with_columns(pl.col("ts").str.to_datetime())<\/p>\n aligned = df.join_asof(events, on="ts", by="device", strategy="backward")<\/p>\n \u6bcf\u4e2a\u70b9\u4f4d\u81ea\u52a8\u5e26\u4e0a\u6700\u8fd1\u4e00\u6b21\u53d1\u751f\u7684\u4e8b\u4ef6，\u505a\u6839\u56e0\u56de\u6eaf\u975e\u5e38\u597d\u7528\u3002<\/p>\n 5.6 \u6d41\u5f0f\u5904\u7406：\u5904\u7406\u8d85\u8fc7\u5185\u5b58\u7684 Parquet<\/h4>\n# \u5047\u8bbe hourly \u76ee\u5f55\u4e0b\u6709 10 GB Parquet \u5206\u7247
\nlf = pl.scan_parquet("data\/hourly\/*.parquet") # scan_* \u7cfb\u5217\u5929\u7136\u60f0\u6027
\nsummary = (
\n lf.filter(pl.col("q") == 1)
\n .group_by(["device", pl.col("ts").dt.truncate("1h")])
\n .agg(pl.col("value").mean())
\n .collect(streaming=True)
\n)<\/p>\n scan_parquet\u00a0\u53ea\u8bfb\u53d6\u67e5\u8be2\u6240\u9700\u7684\u5217（\u5217\u88c1\u526a）+ \u53ea\u8bfb\u53d6\u6ee1\u8db3\u8fc7\u6ee4\u7684\u884c\u7ec4（\u8c13\u8bcd\u4e0b\u63a8），\u914d\u5408\u00a0streaming=True\u00a0\u53ef\u4ee5\u5904\u7406\u8fdc\u8d85\u5185\u5b58\u7684\u6570\u636e\u3002<\/p>\n 5.7 \u4e0e Pandas \/ Matplotlib \u4e92\u64cd\u4f5c<\/h4>\n# Polars \u7b97，Pandas\/Matplotlib \u753b（\u6216\u76f4\u63a5 Matplotlib \u6d88\u8d39 numpy）
\nimport matplotlib.pyplot as plt<\/p>\n trend = (
\n df.lazy()
\n .group_by(pl.col("ts").dt.truncate("1h"))
\n .agg(pl.col("load").mean())
\n .collect()
\n)
\nplt.plot(trend["ts"].to_numpy(), trend["load"].to_numpy())<\/p>\n 5.8 \u751f\u4ea7\u9aa8\u67b6：\u6279\u91cf\u843d\u76d8 + \u589e\u91cf\u8ffd\u52a0 + \u538b\u7f29<\/h4>\n# \u6bcf\u5c0f\u65f6\u628a\u805a\u5408\u7ed3\u679c\u8ffd\u52a0\u5230 Parquet（\u7528 pyarrow \u6216\u76f4\u63a5\u5199\u5206\u7247\u6587\u4ef6）
\ndef write_hourly(df: pl.DataFrame, base: str):
\n hour = df["ts"].min().strftime("%Y%m%d_%H")
\n path = f"{base}\/agg_{hour}.parquet"
\n df.write_parquet(path, compression="zstd", compression_level=19)<\/p>\n
\n6. \u5e95\u5c42\u5b9e\u73b0\u5256\u6790（\u5b83\u4e3a\u4ec0\u4e48\u8fd9\u4e48\u5feb）<\/h3>\n6.1 Apache Arrow \u5217\u5f0f\u5185\u5b58\u683c\u5f0f<\/h4>\nPolars \u7684\u6240\u6709\u6570\u636e\u90fd\u88c5\u5728\u00a0Arrow \u7684 ChunkedArray\u00a0\u91cc\u3002\u6bcf\u5217\u662f\u4e00\u5757\u8fde\u7eed\u7684\u540c\u7c7b\u578b\u7f13\u51b2，\u53e6\u5916\u8fd8\u6709\u4e00\u4e2a validity bitmap（\u4f4d\u56fe）\u8868\u793a\u7a7a\u503c\u2014\u2014\u6240\u4ee5\u7a7a\u503c\u4e0d\u9700\u8981\u5355\u72ec\u7684 NaN \u54e8\u5175，\u4e5f\u4e0d\u5360 8 \u5b57\u8282\u3002\u5185\u5b58\u5e03\u5c40\u5bf9 CPU \u7f13\u5b58\u6781\u5ea6\u53cb\u597d：\u626b\u63cf\u4e00\u5217 = \u987a\u5e8f\u8bfb\u4e00\u7247\u8fde\u7eed\u5185\u5b58\u3002<\/p>\n 6.2 \u591a\u7ebf\u7a0b\u5e76\u884c<\/h4>\nPolars \u9ed8\u8ba4\u4f7f\u7528\u673a\u5668\u5168\u90e8\u6838\u5fc3\u3002\u5b83\u7684\u5e76\u884c\u7b56\u7565\u662f：<\/p>\n \n- \u6309\u5217\u5e76\u884c：select\u00a0\u91cc\u591a\u4e2a\u8868\u8fbe\u5f0f\u53ef\u4ee5\u5e76\u884c\u6c42\u503c；<\/li>\n
- \u6309\u5757\u5e76\u884c：\u5355\u5217\u64cd\u4f5c\u628a\u6570\u636e\u5207\u6210 chunk \u5e76\u884c\u5904\u7406，\u518d\u62fc\u63a5；<\/li>\n
- \u6309\u7ec4\u5e76\u884c：group_by\u00a0\u5148\u5e76\u884c\u54c8\u5e0c\u5206\u7ec4，\u518d\u5408\u5e76\u5404\u7ec4\u7ed3\u679c\u3002<\/li>\n<\/ul>\n
\u5bf9\u6bd4 Pandas \u7684\u5355\u7ebf\u7a0b，8 \u6838\u673a\u5668\u4e0a\u7eaf\u8ba1\u7b97\u4efb\u52a1\u8f7b\u677e\u5feb 5~20 \u500d\u3002<\/p>\n 6.3 SIMD \u5411\u91cf\u5316<\/h4>\nRust \u7f16\u8bd1\u5668（LLVM）\u81ea\u52a8\u4e3a\u70ed\u5faa\u73af\u751f\u6210 AVX2\/AVX-512 \u6307\u4ee4，\u4e00\u6b21\u5904\u7406 8 \u4e2a float64 \u6216 16 \u4e2a int32\u3002\u5217\u5f0f\u5e03\u5c40\u8ba9\u300c\u6574\u5217\u662f\u8fde\u7eed\u540c\u7c7b\u578b\u6570\u7ec4\u300d\u6210\u4e3a\u5e38\u6001，SIMD \u6536\u76ca\u6700\u5927\u5316\u3002<\/p>\n 6.4 \u67e5\u8be2\u4f18\u5316\u5668（Lazy \u7684\u79d8\u5bc6）<\/h4>\nPolars \u7684 LazyFrame \u4f1a\u505a\u4ee5\u4e0b\u4f18\u5316（explain()\u00a0\u53ef\u89c1）：<\/p>\n | | | |