并不是所有的过滤器都能被按下。一般来说，大多数包含函数调用的过滤器 substring 或者 unix_timestamp 不能向下推。在datasourcestrategy中实现了过滤器被下推的完整逻辑。
在这种情况下，解决此限制的一种方法是存储 eventDTLocal 作为unix时间戳而不是parquet文件中的字符串，然后按特定的毫秒进行过滤。

# create some test data

data = [(52.5151923, 13.3824107, 1618760421000), 
        (1.0, 1.0, 1)]
spark.createDataFrame(data, schema=['lat', 'lng', 'eventDTLocal']) \
    .write.mode("overwrite").parquet("dataWithUnixTime")

# get the first and last millisecond of the day

# the timezone has probably to be adjusted

from datetime import datetime, timezone
dt = datetime(2021, 4, 18)
start = dt.replace(tzinfo=timezone.utc).timestamp() * 1000
end = start + 24 * 60 * 60 * 1000 - 1

# run the query

df = spark.read.parquet("dataWithUnixTime") \
    .filter(f"eventDTLocal >= {start} and eventDTLocal <= {end}")

物理平面图 df ```
== Physical Plan ==

• (1) Project [lat#9, lng#10, eventDTLocal#11L]

+- *(1) Filter ((isnotnull(eventDTLocal#11L) AND (eventDTLocal#11L >= 1618704000000)) AND (eventDTLocal#11L <= 1618790399999))
+- *(1) ColumnarToRow
+- FileScan parquet [lat#9,lng#10,eventDTLocal#11L] Batched: true, DataFilters: [isnotnull(eventDTLocal#11L), (eventDTLocal#11L >= 1618704000000), (eventDTLocal#11L <= 161879039..., Format: Parquet, Location: InMemoryFileIndex[file:/home/werner/Java/pyspark3/dataWithUnixTime], PartitionFilters: [], PushedFilters: [IsNotNull(eventDTLocal), GreaterThanOrEqual(eventDTLocal,1618704000000), LessThanOrEqual(eventDT..., ReadSchema: structlat:double,lng:double,eventDTLocal:bigint

现在包含推送的过滤器 `GreaterThanOrEqual` 以及 `LessThanOrEqual` 对于日期列。