Table of Contents |
---|
This page is meant as a template for writing a KIP. To create a KIP choose Tools->Copy on this page and modify with your content and replace the heading with the next KIP number and a description of your issue. Replace anything in italics with your own description.
Status
Current state: "Under Discussion"Accepted
Discussion thread: here [Change the link from the KIP proposal email archive to your own email thread]
Vote thread: here
JIRA: KAFKA-13511
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
Motivation
...
Currently, the Kafka Connect SMT TimestampConverter can convert Timestamp from multiples sources types (String, Unix Long or Date) into different target types (String, Unix Long or Date).
The problem is that Unix Long as a source or as a target type is required to be epoch in millisecondsmust be with milliseconds precision.
In many cases, epoch Unix time is represented with different precisions within external systems : seconds, milliseconds, microseconds, nanoseconds.
When such case arise, Kafka Connect can't do anything expect pass along the Unix Long and leave the conversion to another layer.
This issue was raised several times :
TimestampConverter should have a config to define which precision to use when converting from and to Long Unix timestamp.
Public Interfaces
...
name | description | type | default | valid values | importance |
---|---|---|---|---|---|
unix.precision | The desired unix precision for the timestamp. Used to generate the output when type=unix or used to parse the input if the input is a Long. Note: This SMT will cause precision loss during conversions from and to values with sub-milliseconds components. | String | milliseconds | , , , | low |
Proposed Changes
New config property for TimestampConverter
Implementation details to be discussed :
- TimeUnit.MILLISECONDS.toMicros(unixMillis) and so on for the other conversions seems the easiest way.
Unix Long to Timestamp example:
Code Block | ||
---|---|---|
| ||
"transforms.TimestampConverter.type": "org.apache.kafka.connect.transforms.TimestampConverter$Value", "transforms.TimestampConverter.field": "event_date_long", "transforms.TimestampConverter.epochunix.precision": "microsmicroseconds", "transforms.TimestampConverter.target.type": "Timestamp" |
String to Unix Long nanoseconds example:
Code Block | ||
---|---|---|
| ||
"transforms.TimestampConverter.type": "org.apache.kafka.connect.transforms.TimestampConverter$Value",
"transforms.TimestampConverter.field": "event_date_str",
"transforms.TimestampConverter.format": "yyyy-MM-dd'T'HH:mm:ss.SSS",
"transforms.TimestampConverter.target.type": "unix" |
int32 and seconds
,
"transforms.TimestampConverter.unix.precision": "nanoseconds" |
java.util.Date and SimpleDateFormat limitations
Since these classes can only handle precisions down to the millisecond, it should be noted that:
- converting source Unix Long microseconds or nanos into any target type leads to a precision loss (truncation after millis)
- converting any source type into target Unix Long microseconds or nanos, the part after milliseconds will always be 0
- A KIP that address Date vs Instant may be more appropriate but it impacts so much of the code that I believe this is a good first step.
int32 and seconds
Systems Since int32 should not be used for timestamp, obsolete systems that produces int32 into Kafka and want should willingly chain Cast SMT and then TimestampConverter SMT if they want to use this feature.
Code Block | ||
---|---|---|
| ||
"transforms": "Cast,TimestampConverter", "transforms.Cast.type": "org.apache.kafka.connect.transforms.Cast$Value", "transforms.Cast.spec": "event_date_int:int64", "transforms.TimestampConverter.type": "org.apache.kafka.connect.transforms.TimestampConverter$Value", "transforms.TimestampConverter.field": "event_date_int", "transforms.TimestampConverter.epochunix.precision": "seconds", "transforms.TimestampConverter.target.type": "Timestamp" |
java.util.Date and SimpleDateFormat limitations
...
- converting Long microseconds into any target type leads to a precision loss (truncation after millis)
- converting any source type into Long microseconds, the microseconds part will always be 000
This raises an interesting question (out of the scope of this KIP I guess) regarding the future of java.util.Date
Compatibility, Deprecation, and Migration Plan
The change will not break the compatibility.
Rejected Alternatives
If there are alternative ways of accomplishing the same thing, what were they? The purpose of this section is to motivate why the design is the way it is and not some other way.
Name the field epoch.precision
Since epoch is not a measure but rather a point in time, it can't be associated to a precision.
It makes more sense to name the field unix.precision
for that reason.
Use , , , or as values
seconds is a unit, but millis micros, nanos are really just prefixes. Mixing up doesn't work well.
s, ms, µs, ns are a valid SI symbols but the µs (or its accepted equivalent us) can be confusing.
For clarity, it was decided to use the plaintext naming convention.