Hi all,
I'm +1 with what everything you said Marios.
I'm gonna add another argument on top of that: the "legacy-cast-behavior"
has also a broken type inference, leading to incorrect results or further
errors down in the pipeline[1]. For example, take this:
SELECT COALESCE(CAST('a' AS INT), 0) ...
With the legacy cast behavior ENABLED, this is going to lead to the wrong
result, as 'a' is inferred as VARCHAR NOT NULL, the CAST return value is
inferred as INT NOT NULL, so the planner will drop COALESCE, and will never
return 0. Essentially, CAST will infer the wrong nullability leading to
assigning its result to a NOT NULL type, when its value can effectively be
NULL.
> You introduce a deprecated flag to help users
using old versions of the software to smoothly transition to the new
version, while the new users experience the new features/behavior,
without the need to set a flag.
This is IMO the major point on why we should disable the legacy cast
behavior by default. This is even more relevant with 1.15 and the upgrade
story, as per the problem described above, because users will now end up
generating plans with wrong type inference, which will be hard to migrate
in the next flink versions.
FG
[1] In case you're wondering why it wasn't fixed, the reason is that fixing
it means creating a breaking change, for details
https://github.com/apache/flink/pull/18611#issuecomment-1028174877
On Tue, Feb 22, 2022 at 10:07 AM Marios Trivyzas <mat...@gmail.com> wrote:
> Hello all!
>
> I would like to bring back the discussion regarding the
> *table.exec.legacy-cast-behaviour*
> configuration option which we are introducing with Flink *1.15*. This
> option provides the users
> with the flexibility to continue using the old (incorrect, according to SQL
> standards) behaviour
> of *CAST.*
>
> With Flink *1.15* we have introduced a bunch of fixes, improvements and new
> casting functionality
> between types, see https://issues.apache.org/jira/browse/FLINK-24403, and
> some of them are
> guarded behind the legacy behaviour option:
>
> - Trimming and padding when casting to CHAR/VARCHAR types to respect the
> specified length
> - Changes for casting various types to CHAR/VARCHAR/STRING
> - Runtime errors for CAST no longer emit *null *as result but exceptions
> are thrown with a meaningful message for the cause, that fail the
> pipeline. *TRY_CAST
> *is introduced instead, which emits *null* results instead of throwing
> exceptions.
>
> Those changes become active if users set the
> *table.exec.legacy-cast-behaviour
> *option to *DISABLED*, otherwise
> they will continue to experience the old, *erroneous*, behaviour of *CAST*.
>
> Currently, we have set the *table.exec.legacy-cast-behaviour *option
> to be *ENABLED
> *by default, so if users want
> to get the new correct behaviour, they are required to set explicitly the
> option to *DISABLED*. Moreover, the option
> itself is marked as deprecated, since the plan is to be removed in the
> future, so that the old, erroneous behaviour
> won't be an option, and the *TRY_CAST* would be the way to go if users
> don't want to have errors and failed pipelines,
> but have *null*s emitted in case of runtime errors when casting.
>
> I would like to start a discussion and maybe ask for voting, so that we set
> the *table.exec.legacy-cast-behaviour* option
> to *DISABLED *by default, so that users that want to keep their old
> pipelines working the same way, without changing their
> SQL/TableAPI code, would need to explicitly set it to *ENABLED.*
>
> My main argument for changing the default value for the option, is that the
> *DISABLED* value is the one that enables
> the *correct* behaviour for CAST which should be the default for all new
> users. This way, new FLINK users, or users which
> build new pipelines, from now on would get the correct behaviour by default
> without the need of changing some flag in their
> configuration. It feels weird to me, especially for people very familiar
> with standard SQL, to be obliged to set some config
> flag, to be able to get the correct behaviour for CAST. On top, users that
> won't read about this option in our docs, will,
> "blindly", experience the old incorrect behaviour for their new pipelines,
> and issues that could cause the CAST to fail
> will remain hidden from them, since *nulls* would be emitted. IMHO, this
> last part is also very important during the development
> stages of an application/pipeline. Normally, developers would want to see
> all possible errors/scenarios during development
> stages , in order to build a robust production system. If errors, are
> hidden then, they can easily end up with those errors
> in the production system which would be even harder to discover and debug,
> since no exception will ever be thrown.
> Imagine that there is a CAST which generates nulls because of runtime
> errors, and its result is used in an aggregate function:
>
> SELECT AVG(CAST(col1 AS FLOAT)), ... FROM t GROUP BY ....
>
> If nulls are emitted by the CAST (let's say because some records have a
> quoted string value for col1, then simply the AVG result
> would be wrong and in would be extremely difficult to realise and fix the
> issue by simply wrapping col1 first with, for example,
> a REGEXP_REPLACE, to get rid of the quotes.
>
> My second argument is that, since we have marked
> *table.exec.legacy-cast-behaviour* as deprecated, and we want to completely
> remove it in the future, if the default value is *DISABLED*, when it's
> removed we also make a breaking change, by changing the
> default behaviour for all users, which is against the common software
> practices. You introduce a deprecated flag to help users
> using old versions of the software to smoothly transition to the new
> version, while the new users experience the new features/behaviour,
> without the need to set a flag. So, when in the future this flag is
> completely removed, for those "new" users it would be completely
> transparent. If we continue with having the default value *ENABLED*, those
> new user would experience an "unnatural" breaking change
> when this option is completely removed.
>
> Best,
> Marios
>
