Core: Fix numeric overflow of timestamp nano literal #11775
Conversation
ebyhr
force-pushed
the
ebi/timestamp-ns
branch
from
5228ca6 to
efe1d14
Compare
| @@ -300,8 +300,7 @@ public <T> Literal<T> to(Type type) { | |||
| case TIMESTAMP: | |||
| return (Literal<T>) new TimestampLiteral(value()); | |||
| case TIMESTAMP_NANO: | |||
| // assume micros and convert to nanos to match the behavior in the timestamp case above | |||
| return new TimestampLiteral(value()).to(type); | |||
| return (Literal<T>) new TimestampNanoLiteral(value()); | |||
There was a problem hiding this comment.
This change seems correct to me, the previous behavior for this was to assume the value was in microseconds and then pass that through to TimestampLiteral but that can overflow and does not actually represent a nanosecond timestamp!
There was a problem hiding this comment.
FWIW I still think this is correct but it's worth getting other's perspective on this since we are changing one of the assumptions of how value is interpreted when the type to convert to is nanoseconds.
CC @nastra @epgif @jacobmarble @rdblue thoughts?
There was a problem hiding this comment.
I think both before and after this change is correct. In Iceberg we had the assumption that everything is in microseconds. But this doesn't hold anymore now we have nano's. I do think the version after the change is more correct and more closely aligns with my expectations. If we can make sure that folks are not using this yet, I think this change is a good one 👍
There was a problem hiding this comment.
I had a chat with @rdblue who reviewed the PR that introduced this, and it is actually on purpose. Spark always passes in microseconds, changing this would break this assumption with Spark. So I think we have to revert this line. That said, I do think we need to check (and raise an error) when it overflows. Easiest way of doing this is by converting it to nano's, and convert is back to micro's and check if it still the same value.
There was a problem hiding this comment.
Thank you for sharing the context. What about other query engines? Actually, I found this issue when I was trying to support nanosecond precision in Trino Iceberg connector. As you may know, the max precision in Trino is picos (12).
| assertThat(Literal.of(400000L).to(TimestampNanoType.withoutZone()).toByteBuffer().array()) | ||
| .isEqualTo(new byte[] {0, -124, -41, 23, 0, 0, 0, 0}); | ||
| .isEqualTo(new byte[] {-128, 26, 6, 0, 0, 0, 0, 0}); | ||
| assertThat(Literal.of(400000L).to(TimestampNanoType.withZone()).toByteBuffer().array()) |
There was a problem hiding this comment.
I'm a bit confused how the original assertion was passing? Shouldn't have this always been equivalent to {-128, 26, 6, 0, 0, 0, 0, 0}?
There was a problem hiding this comment.
I think the cause is the original logic called DateTimeUtil.microsToNanos method which multiples the value by 1000:
iceberg/api/src/main/java/org/apache/iceberg/expressions/Literals.java
Lines 444 to 445 in b9b61b1
There was a problem hiding this comment.
Ah I see the comment on line 107/108. Could we update the assertConversion to instead test against 400000L and then remove the comment. At this point we are no longer having to pass in different values since we Literal.of(someLong).to(TimestampNanos) will always interpret someLong as nanoseconds.
api/src/test/java/org/apache/iceberg/expressions/TestTimestampLiteralConversions.java
Show resolved
Hide resolved
The previous logic leads to overflow when we pass timestamp nanos long value.