Uh oh!
There was an error while loading. Please reload this page.
Support Timestamp type in xlang JDBC Read and Write - #22561
Conversation
3dc1a57 to
006885eCompareAbacn
commented
Aug 2, 2022
Run Python 3.8 PostCommit |
661be51 to
77a42c6CompareAbacn
commented
Aug 2, 2022
Run Python 3.8 PostCommit |
Abacn
commented
Aug 3, 2022
Run Python 3.8 PostCommit |
Codecov Report
@@ Coverage Diff @@## master #22561 +/- ##
==========================================
+ Coverage 73.69% 73.73% +0.04%
==========================================
Files 713 713 Lines 94821 94974 +153 ==========================================
+ Hits 69874 70030 +156 + Misses 23648 23645 -3
Partials 1299 1299
Flags with carried forward coverage won't be shown. Click here to find out more.
📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more |
dd7cb95 to
aec0fd8CompareAbacn
commented
Aug 3, 2022
Run Python 3.8 PostCommit |
Abacn
commented
Aug 3, 2022
Run Python 3.8 PostCommit |
Abacn
commented
Aug 3, 2022
Stopping reviewer notifications for this pull request: review requested by someone other than the bot, ceding control |
c8647eb to
8d4575aCompareAbacn
commented
Aug 3, 2022
Rebased in order to update CHANGES.md. ready for review now. |
Abacn
commented
Aug 3, 2022
Run Python 3.8 PostCommit |
There was a problem hiding this comment.
Do we need this? I thought the solution here was to use the MicrosInstant logical type
There was a problem hiding this comment.
You are right. removed this class the test still pass and confirmed with database content.
There was a problem hiding this comment.
ah messed up local virtualenv. Remove @LogicalType.register_logical_type or the entire class here causes the following error:
Traceback (most recent call last):
File "...virtualenv/py38beam/lib/python3.8/site-packages/apache_beam/typehints/schemas.py", line 408, in typing_from_runner_api
field_py_type = self.typing_from_runner_api(field.type)
File "...virtualenv/py38beam/lib/python3.8/site-packages/apache_beam/typehints/schemas.py", line 362, in typing_from_runner_api
base = self.typing_from_runner_api(base_type)
File "...virtualenv/py38beam/lib/python3.8/site-packages/apache_beam/typehints/schemas.py", line 447, in typing_from_runner_api
return LogicalType.from_runner_api(
File "...virtualenv/py38beam/lib/python3.8/site-packages/apache_beam/typehints/schemas.py", line 621, in from_runner_api
raise ValueError(
ValueError: No logical type registered for URN 'beam:logical_type:datetime:v1'
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
...
File "...virtualenv/py38beam/lib/python3.8/site-packages/apache_beam/typehints/schemas.py", line 179, in typing_from_runner_api
return SchemaTranslation(
File "...virtualenv/py38beam/lib/python3.8/site-packages/apache_beam/typehints/schemas.py", line 412, in typing_from_runner_api
raise ValueError(
ValueError: Failed to decode schema due to an issue with Field proto:
name: "f_timestamp"
type {
nullable: true
logical_type {
urn: "beam:logical_type:datetime:v1"
representation {
atomic_type: INT64
}
}
}
id: 1
encoding_position: 1
Still need this to register the URN.
There was a problem hiding this comment.
I'm confused though, how did we get a field with URN beam:logical_type:datetime:v1? Based on the change in JdbcUtil I thought the intention here was to map JDBC timestamp columns to the MicrosInstant logical type, with URN beam:logical_type:micros_instant:v1, in order to avoid the proiblematic InstantCoder in DATETIME.
There was a problem hiding this comment.
The change in JdbcUtil and Python are for different purposes. Basically there are two things missing for java<->portable timestamp: (1) Java JDBCIO support of portable timestamp; (2) Python support for java timestamp.
The JdbcUtil change is for write to jdbc. It adds support in Java's JDBCIO of converting portable timestamp type (beam:logical_type:micros_instant:v1) to sql timestamp;
The Python coder and schema change is for read from jdbc. It adds supports in Python's RowCoder of identifying (by schema change) and decoding (by coder change) the java timestamp type (datetime) to language timestamp. The field 'beam:logical_type:datetime:v1' is passed from java here when the row contains datetime field type.
There was a problem hiding this comment.
I tried to change coder for Instant in java side but it did not quite work (see comment here) and ended up adding support for InstantCoder coded timestamp in Python side. I understand this is less elegant but it does not touch java Instant coders which is used everywhere not only in schema.
There was a problem hiding this comment.
Can we make portable timestamp and java timestamp the same type? I think it would be preferable to produce micros_instant for the timestamp type in the JDBC read. I'd really like to avoid having Java's DATETIME primitive leak into Python.
There was a problem hiding this comment.
I agree this is preferable in terms of consistency. To do this we will need to make distinction of timestamp in schema based IOs (including JDBC) and windowing timestamp, making the former produce micros_instant instead of joda.Instant, which is similar to the approach here #11456. However given the scope of breaking changes this may leave as next major version change...
If I understood correctly it's not leaking Java's DATETIME primitive (Schema.FieldType DATETIME), the schema passed to python is the logical type beam:logical_type:datetime:v1 that not yet implemented in python. However the coder that used to decode this type is actually available in python which is TimestampCoder. This is still in line with Portable Beam Schema design doc.
There was a problem hiding this comment.
Could there be a new logical type then for use in JDBC read and write that uses millisecond precision, and maps to joda.Instant?
There was a problem hiding this comment.
Now renamed this usage to millis_instant logical type
There was a problem hiding this comment.
I wish we had a more general solution for mapping these to java time vs. joda time
There was a problem hiding this comment.
Here is a mapping from java.time.Instant to java.sql.Timestamp, joda.time is not involved. The comment was a clarification because I see joda.time.Instant is usually used in the codebase (that's why I did not declare import Instant but write down the whole class name here)
Do you think we need a general Row.getJavaInstant() as Row.getDateTime() ?
There was a problem hiding this comment.
Ah ok, my mistake. Regardless there's no action needed here, I was just lamenting that this date/time type situation is messy.
There was a problem hiding this comment.
If we do implement beam:logical_type:datetime in Python we should:
- Add the logical type in schema.proto, and reference the URN from that file:
- Consider if we can keep the logic contained in the logical type implementation, e.g. can we do the byte flipping in a to_language_type and to_representation_type implementation?
- Add some test cases to standard_coders.yaml, like this one: You might also need to skip this case in Go if the "logical" filter doesn't catch it:
There was a problem hiding this comment.
Thanks for the instructions! Will complete the proto change. beam:logical_type:datetime:v1 is the millisecond precision timestamp backed by a fixed length INT64 that encoded with BigEndianLong. Will complete the proto change.
can we do the byte flipping in a to_language_type and to_representation_type implementation?
Considered this before. Under the current framework the value is already decoded with VarInt coder before sent to to_language_type or to_representation_type which is incorrect. All I need is a 8-byte fixed long integer primitive which does not exist in portable primitives, and there is not yet a consensus on reference coder (context). Therefore I ended up this solution to make things work while not introducing new primitives or other fundamental changes.
There was a problem hiding this comment.
ah... test case in standard_coders.yaml won't work either because Both micros_instant and millis_instant has Timestamp language type and the former will take over the priority of millis_instant (by design). When a timestamp is decoded to an integer, the test will call micros_instant's to_language_type.
i.e. MillisInstant->Timestamp conversion is one direction and Timestamp <-> MicrosInstant is bidirectional.
There was a problem hiding this comment.
Hmm I'd really like to have this tested at the standard_coders.yaml level as a unit test, it's not ideal if the only thing verifying compatibility is an integration test in Python.
If you have time tomorrow maybe we could do a quick video call to live debug what's going wrong here, and see if we can work around it. If it will be painful we can leave it as future work, but I'd like to understand the level of effort.
There was a problem hiding this comment.
Managed to get a test case using the following code snippet
defgenerate_millis():
# Logical type that deals with millis_instant urn (MillisInstant)MillisLogicalType=LogicalType._known_logical_types.get_logical_type_by_urn('beam:logical_type:millis_instant:v1')
# Original Logical type used to represent Timestamp (MicrosInstant)TimestampLogicalType=LogicalType._known_logical_types.get_logical_type_by_language_type(Timestamp)
LogicalType._known_logical_types.by_language_type[Timestamp] =MillisLogicalTypeschema=beam.typehints.schemas.named_tuple_to_schema(TestTuple)
coder=beam.coders.row_coder.RowCoder(schema)
print("payload = %s"%schema.SerializeToString())
examples= (TestTuple(
f_timestamp=Timestamp.from_rfc3339("2020-08-13T14:14:14.123Z"),
f_string="2020-08-13T14:14:14.123Z",
f_int=1597328054123),)
forexampleinexamples:
print("example = %s"%coder.encode(example))
# recover original registrationLogicalType._known_logical_types.by_language_type[Timestamp] =TimestampLogicalTypeThe workaround is temporarily change the mapping of Timestamp -> MillisInstant logical type. Without it Timestamp always maps to MicrosInstant logical type in Python.
There was a problem hiding this comment.
Ahhh ok, so it looks like the issue was using Python to generate the test values. Another strategy might be to create the schema proto directly. But this way works too and is probably less code, thanks!
d0498fa to
df31d95CompareUh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
There was a problem hiding this comment.
Hmm I'd really like to have this tested at the standard_coders.yaml level as a unit test, it's not ideal if the only thing verifying compatibility is an integration test in Python.
If you have time tomorrow maybe we could do a quick video call to live debug what's going wrong here, and see if we can work around it. If it will be painful we can leave it as future work, but I'd like to understand the level of effort.
dbf0c2f to
5ac3b4aCompareAbacn
commented
Aug 25, 2022
Run Python_PVR_Flink PreCommit |
Abacn
commented
Aug 25, 2022
Run Java PreCommit |
Abacn
commented
Aug 25, 2022
Run Python 3.8 PostCommit |
Abacn
commented
Aug 25, 2022
Run Java PreCommit |
postcommit test Apparently the branch itself works but now failing when merge onto master. Investigating. Update: related to the change of #22679. After it tests fail. |
* Support beam:logical_type:datetime:v1 in python sdk * Support micro_instant logical type in java jdbc * Fix test timestamp milli precision support in mysql
Co-authored-by: Brian Hulette <hulettbh@gmail.com>
a6f94e8 to
b0484e7CompareAbacn
commented
Aug 26, 2022
Rebased onto latest master; the last commit is a fix of test |
Abacn
commented
Aug 26, 2022
Run Python 3.8 PostCommit |
| def register_logical_type(cls, logical_type_cls): | ||
| """Register an implementation of LogicalType.""" | ||
| cls._known_logical_types.add(logical_type_cls.urn(), logical_type_cls) | ||
| return logical_type_cls |
There was a problem hiding this comment.
These places were some minor bugs. If not return logical_type_cls, class decorated with @register_logical_type cannot be imported (will be None) and docstring is also missing from pydoc; signatures of to_representation_type and to_language_type should have a value parameter as its child classes. Fixed here.
Abacn
commented
Aug 27, 2022
Run Java PreCommit |
Abacn
commented
Aug 27, 2022
Run Python PreCommit |
1 similar comment
Abacn
commented
Aug 29, 2022
Run Python PreCommit |
Part of #19817
Please add a meaningful description for your change here
Thank you for your contribution! Follow this checklist to help us incorporate your contribution quickly and easily:
R: @username).addresses #123), if applicable. This will automatically add a link to the pull request in the issue. If you would like the issue to automatically close on merging the pull request, commentfixes #<ISSUE NUMBER>instead.CHANGES.mdwith noteworthy changes.See the Contributor Guide for more tips on how to make review process smoother.
To check the build health, please visit https://github.com/apache/beam/blob/master/.test-infra/BUILD_STATUS.md
GitHub Actions Tests Status (on master branch)
See CI.md for more information about GitHub Actions CI.