spark-avro Read/write timestamps

Reading/writing dataframes to/from avro files with timestamps with spark gives an inconsistent behavior. The point being that the conversion between timestamps and integers has a different logic in spark vs. what happens working with avro: spark-avro converts timestamps to long int in milliseconds; however, casting them back to "timestamp" in Spark gives wildly wrong dates, as Spark expects timestamps in seconds, I believe (dividing by 1000 after you read the data is an obvious workaraound).

Example:

scala> val df = Seq("2017-05-05", "2016-04-02").toDF.select($"value".cast("timestamp").alias("time"))

scala> df.printSchema()
root
|-- ts: timestamp (nullable = true)

scala> df.show(truncate=false)
+---------------------+
|ts                   |
+---------------------+
|2017-05-05 00:00:00.0|
|2016-04-02 00:00:00.0|
+---------------------+

scala> df.avro.write("test.avro")

scala> spark.read.avro("test.avro").show()
+-------------+
|           ts|
+-------------+
|1493935200000|
|1459548000000|
+-------------+

scala> spark.read.avro("test.avro").select($"ts".cast("timestamp")).show(truncate=false)
+----------------------+
|                    ts|
+----------------------+
|49310-12-03 17:00:00.0|
|48221-03-27 18:00:00.0|
+----------------------+

scala> spark.read.avro("test.avro").select(($"ts"/1000).cast("timestamp")).show(truncate=false)
+------------------------------+
|CAST((ts / 1000) AS TIMESTAMP)|
+------------------------------+
|2017-05-05 00:00:00.0         |
|2016-04-02 00:00:00.0         |
+------------------------------+

I'm wondering if there's any better way to handle this behavior than having to manually divide the timestamp fields by 1000 when you read from avro.