[VOTE][SPARK-28885] Follow ANSI store assignment rules in table insertion by default

classic Classic list List threaded Threaded
8 messages Options
Reply | Threaded
Open this post in threaded view
|

[VOTE][SPARK-28885] Follow ANSI store assignment rules in table insertion by default

Gengliang Wang
Hi everyone,

I'd like to call for a vote on SPARK-28885 "Follow ANSI store assignment rules in table insertion by default".  
When inserting a value into a column with the different data type, Spark performs type coercion. Currently, we support 3 policies for the type coercion rules: ANSI, legacy and strict, which can be set via the option "spark.sql.storeAssignmentPolicy":
1. ANSI: Spark performs the type coercion as per ANSI SQL. In practice, the behavior is mostly the same as PostgreSQL. It disallows certain unreasonable type conversions such as converting `string` to `int` and `double` to `boolean`.
2. Legacy: Spark allows the type coercion as long as it is a valid `Cast`, which is very loose. E.g., converting either `string` to `int` or `double` to `boolean` is allowed. It is the current behavior in Spark 2.x for compatibility with Hive.
3. Strict: Spark doesn't allow any possible precision loss or data truncation in type coercion, e.g., converting either `double` to `int` or `decimal` to `double` is allowed. The rules are originally for Dataset encoder. As far as I know, no maintainstream DBMS is using this policy by default.

Currently, the V1 data source uses "Legacy" policy by default, while V2 uses "Strict". This proposal is to use "ANSI" policy by default for both V1 and V2 in Spark 3.0.

There was also a DISCUSS thread "Follow ANSI SQL on table insertion" in the dev mailing list.

This vote is open until next Thurs (Sept. 12nd).

[ ] +1: Accept the proposal
[ ] +0
[ ] -1: I don't think this is a good idea because ...
Thank you!
Gengliang
Reply | Threaded
Open this post in threaded view
|

Re: [VOTE][SPARK-28885] Follow ANSI store assignment rules in table insertion by default

cloud0fan
+1

To be honest I don't like the legacy policy. It's too loose and easy for users to make mistakes, especially when Spark returns null if a function hit errors like overflow.

The strict policy is not good either. It's too strict and stops valid use cases like writing timestamp values to a date type column. Users do expect truncation to happen without adding cast manually in this case. It's also weird to use a spark specific policy that no other database is using.

The ANSI policy is better. It stops invalid use cases like writing string values to an int type column, while keeping valid use cases like timestamp -> date.

I think it's no doubt that we should use ANSI policy instead of legacy policy for v1 tables. Except for backward compatibility, ANSI policy is literally better than the legacy policy.

The v2 table is arguable here. Although the ANSI policy is better than strict policy to me, this is just the store assignment policy, which only partially controls the table insertion behavior. With Spark's "return null on error" behavior, the table insertion is more likely to insert invalid null values with the ANSI policy compared to the strict policy.

I think we should use ANSI policy by default for both v1 and v2 tables, because
1. End-users don't care how the table is implemented. Spark should provide consistent table insertion behavior between v1 and v2 tables.
2. Data Source V2 is unstable in Spark 2.x so there is no backward compatibility issue. That said, the baseline to judge which policy is better should be the table insertion behavior in Spark 2.x, which is the legacy policy + "return null on error". ANSI policy is better than the baseline.
3. We expect more and more uses to migrate their data sources to the V2 API. The strict policy can be a stopper as it's a too big breaking change, which may break many existing queries.

Thanks,
Wenchen 


On Wed, Sep 4, 2019 at 1:59 PM Gengliang Wang <[hidden email]> wrote:
Hi everyone,

I'd like to call for a vote on SPARK-28885 "Follow ANSI store assignment rules in table insertion by default".  
When inserting a value into a column with the different data type, Spark performs type coercion. Currently, we support 3 policies for the type coercion rules: ANSI, legacy and strict, which can be set via the option "spark.sql.storeAssignmentPolicy":
1. ANSI: Spark performs the type coercion as per ANSI SQL. In practice, the behavior is mostly the same as PostgreSQL. It disallows certain unreasonable type conversions such as converting `string` to `int` and `double` to `boolean`.
2. Legacy: Spark allows the type coercion as long as it is a valid `Cast`, which is very loose. E.g., converting either `string` to `int` or `double` to `boolean` is allowed. It is the current behavior in Spark 2.x for compatibility with Hive.
3. Strict: Spark doesn't allow any possible precision loss or data truncation in type coercion, e.g., converting either `double` to `int` or `decimal` to `double` is allowed. The rules are originally for Dataset encoder. As far as I know, no maintainstream DBMS is using this policy by default.

Currently, the V1 data source uses "Legacy" policy by default, while V2 uses "Strict". This proposal is to use "ANSI" policy by default for both V1 and V2 in Spark 3.0.

There was also a DISCUSS thread "Follow ANSI SQL on table insertion" in the dev mailing list.

This vote is open until next Thurs (Sept. 12nd).

[ ] +1: Accept the proposal
[ ] +0
[ ] -1: I don't think this is a good idea because ...
Thank you!
Gengliang
Reply | Threaded
Open this post in threaded view
|

Re: [VOTE][SPARK-28885] Follow ANSI store assignment rules in table insertion by default

rxin
Having three modes is a lot. Why not just use ansi mode as default, and legacy for backward compatibility? Then over time there's only the ANSI mode, which is standard compliant and easy to understand. We also don't need to invent a standard just for Spark.


On Thu, Sep 05, 2019 at 12:27 AM, Wenchen Fan <[hidden email]> wrote:
+1

To be honest I don't like the legacy policy. It's too loose and easy for users to make mistakes, especially when Spark returns null if a function hit errors like overflow.

The strict policy is not good either. It's too strict and stops valid use cases like writing timestamp values to a date type column. Users do expect truncation to happen without adding cast manually in this case. It's also weird to use a spark specific policy that no other database is using.

The ANSI policy is better. It stops invalid use cases like writing string values to an int type column, while keeping valid use cases like timestamp -> date.

I think it's no doubt that we should use ANSI policy instead of legacy policy for v1 tables. Except for backward compatibility, ANSI policy is literally better than the legacy policy.

The v2 table is arguable here. Although the ANSI policy is better than strict policy to me, this is just the store assignment policy, which only partially controls the table insertion behavior. With Spark's "return null on error" behavior, the table insertion is more likely to insert invalid null values with the ANSI policy compared to the strict policy.

I think we should use ANSI policy by default for both v1 and v2 tables, because
1. End-users don't care how the table is implemented. Spark should provide consistent table insertion behavior between v1 and v2 tables.
2. Data Source V2 is unstable in Spark 2.x so there is no backward compatibility issue. That said, the baseline to judge which policy is better should be the table insertion behavior in Spark 2.x, which is the legacy policy + "return null on error". ANSI policy is better than the baseline.
3. We expect more and more uses to migrate their data sources to the V2 API. The strict policy can be a stopper as it's a too big breaking change, which may break many existing queries.

Thanks,
Wenchen 


On Wed, Sep 4, 2019 at 1:59 PM Gengliang Wang <[hidden email]> wrote:
Hi everyone,

I'd like to call for a vote on SPARK-28885 "Follow ANSI store assignment rules in table insertion by default".  
When inserting a value into a column with the different data type, Spark performs type coercion. Currently, we support 3 policies for the type coercion rules: ANSI, legacy and strict, which can be set via the option "spark.sql.storeAssignmentPolicy":
1. ANSI: Spark performs the type coercion as per ANSI SQL. In practice, the behavior is mostly the same as PostgreSQL. It disallows certain unreasonable type conversions such as converting `string` to `int` and `double` to `boolean`.
2. Legacy: Spark allows the type coercion as long as it is a valid `Cast`, which is very loose. E.g., converting either `string` to `int` or `double` to `boolean` is allowed. It is the current behavior in Spark 2.x for compatibility with Hive.
3. Strict: Spark doesn't allow any possible precision loss or data truncation in type coercion, e.g., converting either `double` to `int` or `decimal` to `double` is allowed. The rules are originally for Dataset encoder. As far as I know, no maintainstream DBMS is using this policy by default.

Currently, the V1 data source uses "Legacy" policy by default, while V2 uses "Strict". This proposal is to use "ANSI" policy by default for both V1 and V2 in Spark 3.0.

There was also a DISCUSS thread "Follow ANSI SQL on table insertion" in the dev mailing list.

This vote is open until next Thurs (Sept. 12nd).

[ ] +1: Accept the proposal
[ ] +0
[ ] -1: I don't think this is a good idea because ...
Thank you!
Gengliang

Reply | Threaded
Open this post in threaded view
|

Re: [VOTE][SPARK-28885] Follow ANSI store assignment rules in table insertion by default

Alastair Green
Makes sense.

While the ISO SQL standard automatically becomes an American national  (ANSI) standard, changes are only made to the International (ISO/IEC) Standard, which is the authoritative specification. 

These rules are specified in SQL/Foundation (ISO/IEC SQL Part 2), section 9.2. 

Could we rename the proposed default to “ISO/IEC (ANSI)”? 

— Alastair

On Thu, Sep 5, 2019 at 17:17, Reynold Xin <[hidden email]> wrote:
Having three modes is a lot. Why not just use ansi mode as default, and legacy for backward compatibility? Then over time there's only the ANSI mode, which is standard compliant and easy to understand. We also don't need to invent a standard just for Spark.


On Thu, Sep 05, 2019 at 12:27 AM, Wenchen Fan <[hidden email]> wrote:
+1

To be honest I don't like the legacy policy. It's too loose and easy for users to make mistakes, especially when Spark returns null if a function hit errors like overflow.

The strict policy is not good either. It's too strict and stops valid use cases like writing timestamp values to a date type column. Users do expect truncation to happen without adding cast manually in this case. It's also weird to use a spark specific policy that no other database is using.

The ANSI policy is better. It stops invalid use cases like writing string values to an int type column, while keeping valid use cases like timestamp -> date.

I think it's no doubt that we should use ANSI policy instead of legacy policy for v1 tables. Except for backward compatibility, ANSI policy is literally better than the legacy policy.

The v2 table is arguable here. Although the ANSI policy is better than strict policy to me, this is just the store assignment policy, which only partially controls the table insertion behavior. With Spark's "return null on error" behavior, the table insertion is more likely to insert invalid null values with the ANSI policy compared to the strict policy.

I think we should use ANSI policy by default for both v1 and v2 tables, because
1. End-users don't care how the table is implemented. Spark should provide consistent table insertion behavior between v1 and v2 tables.
2. Data Source V2 is unstable in Spark 2.x so there is no backward compatibility issue. That said, the baseline to judge which policy is better should be the table insertion behavior in Spark 2.x, which is the legacy policy + "return null on error". ANSI policy is better than the baseline.
3. We expect more and more uses to migrate their data sources to the V2 API. The strict policy can be a stopper as it's a too big breaking change, which may break many existing queries.

Thanks,
Wenchen


On Wed, Sep 4, 2019 at 1:59 PM Gengliang Wang <[hidden email]> wrote:
Hi everyone,

I'd like to call for a vote on SPARK-28885 "Follow ANSI store assignment rules in table insertion by default".
When inserting a value into a column with the different data type, Spark performs type coercion. Currently, we support 3 policies for the type coercion rules: ANSI, legacy and strict, which can be set via the option "spark.sql.storeAssignmentPolicy":
1. ANSI: Spark performs the type coercion as per ANSI SQL. In practice, the behavior is mostly the same as PostgreSQL. It disallows certain unreasonable type conversions such as converting `string` to `int` and `double` to `boolean`.
2. Legacy: Spark allows the type coercion as long as it is a valid `Cast`, which is very loose. E.g., converting either `string` to `int` or `double` to `boolean` is allowed. It is the current behavior in Spark 2.x for compatibility with Hive.
3. Strict: Spark doesn't allow any possible precision loss or data truncation in type coercion, e.g., converting either `double` to `int` or `decimal` to `double` is allowed. The rules are originally for Dataset encoder. As far as I know, no maintainstream DBMS is using this policy by default.

Currently, the V1 data source uses "Legacy" policy by default, while V2 uses "Strict". This proposal is to use "ANSI" policy by default for both V1 and V2 in Spark 3.0.

There was also a DISCUSS thread "Follow ANSI SQL on table insertion" in the dev mailing list.

This vote is open until next Thurs (Sept. 12nd).

[ ] +1: Accept the proposal
[ ] +0
[ ] -1: I don't think this is a good idea because ...
Thank you!
Gengliang

Reply | Threaded
Open this post in threaded view
|

Re: [VOTE][SPARK-28885] Follow ANSI store assignment rules in table insertion by default

Ryan Blue

We discussed this thread quite a bit in the DSv2 sync up and Russell brought up a really good point about this.

The ANSI rule used here specifies how to store a specific value, V, so this is a runtime rule — an earlier case covers when V is NULL, so it is definitely referring to a specific value. The rule requires that if the type doesn’t match or if the value cannot be truncated, an exception is thrown for “numeric value out of range”.

That runtime error guarantees that even though the cast is introduced at analysis time, unexpected NULL values aren’t inserted into a table in place of data values that are out of range. Unexpected NULL values are the problem that was concerning to many of us in the discussion thread, but it turns out that real ANSI behavior doesn’t have the problem. (In the sync, we validated this by checking Postgres and MySQL behavior, too.)

In Spark, the runtime check is a separate configuration property from this one, but in order to actually implement ANSI semantics, both need to be set. So I think it makes sense to change both defaults to be ANSI. The analysis check alone does not implement the ANSI standard.

In the sync, we also agreed that it makes sense to be able to turn off the runtime check in order to avoid job failures. Another, safer way to avoid job failures is to require an explicit cast, i.e., strict mode.

I think that we should amend this proposal to change the default for both the runtime check and the analysis check to ANSI.

As this stands now, I vote -1. But I would support this if the vote were to set both runtime and analysis checks to ANSI mode.

rb


On Fri, Sep 6, 2019 at 3:12 AM Alastair Green <[hidden email]> wrote:
Makes sense.

While the ISO SQL standard automatically becomes an American national  (ANSI) standard, changes are only made to the International (ISO/IEC) Standard, which is the authoritative specification. 

These rules are specified in SQL/Foundation (ISO/IEC SQL Part 2), section 9.2. 

Could we rename the proposed default to “ISO/IEC (ANSI)”? 

— Alastair

On Thu, Sep 5, 2019 at 17:17, Reynold Xin <[hidden email]> wrote:
Having three modes is a lot. Why not just use ansi mode as default, and legacy for backward compatibility? Then over time there's only the ANSI mode, which is standard compliant and easy to understand. We also don't need to invent a standard just for Spark.


On Thu, Sep 05, 2019 at 12:27 AM, Wenchen Fan <[hidden email]> wrote:
+1

To be honest I don't like the legacy policy. It's too loose and easy for users to make mistakes, especially when Spark returns null if a function hit errors like overflow.

The strict policy is not good either. It's too strict and stops valid use cases like writing timestamp values to a date type column. Users do expect truncation to happen without adding cast manually in this case. It's also weird to use a spark specific policy that no other database is using.

The ANSI policy is better. It stops invalid use cases like writing string values to an int type column, while keeping valid use cases like timestamp -> date.

I think it's no doubt that we should use ANSI policy instead of legacy policy for v1 tables. Except for backward compatibility, ANSI policy is literally better than the legacy policy.

The v2 table is arguable here. Although the ANSI policy is better than strict policy to me, this is just the store assignment policy, which only partially controls the table insertion behavior. With Spark's "return null on error" behavior, the table insertion is more likely to insert invalid null values with the ANSI policy compared to the strict policy.

I think we should use ANSI policy by default for both v1 and v2 tables, because
1. End-users don't care how the table is implemented. Spark should provide consistent table insertion behavior between v1 and v2 tables.
2. Data Source V2 is unstable in Spark 2.x so there is no backward compatibility issue. That said, the baseline to judge which policy is better should be the table insertion behavior in Spark 2.x, which is the legacy policy + "return null on error". ANSI policy is better than the baseline.
3. We expect more and more uses to migrate their data sources to the V2 API. The strict policy can be a stopper as it's a too big breaking change, which may break many existing queries.

Thanks,
Wenchen


On Wed, Sep 4, 2019 at 1:59 PM Gengliang Wang <[hidden email]> wrote:
Hi everyone,

I'd like to call for a vote on SPARK-28885 "Follow ANSI store assignment rules in table insertion by default".
When inserting a value into a column with the different data type, Spark performs type coercion. Currently, we support 3 policies for the type coercion rules: ANSI, legacy and strict, which can be set via the option "spark.sql.storeAssignmentPolicy":
1. ANSI: Spark performs the type coercion as per ANSI SQL. In practice, the behavior is mostly the same as PostgreSQL. It disallows certain unreasonable type conversions such as converting `string` to `int` and `double` to `boolean`.
2. Legacy: Spark allows the type coercion as long as it is a valid `Cast`, which is very loose. E.g., converting either `string` to `int` or `double` to `boolean` is allowed. It is the current behavior in Spark 2.x for compatibility with Hive.
3. Strict: Spark doesn't allow any possible precision loss or data truncation in type coercion, e.g., converting either `double` to `int` or `decimal` to `double` is allowed. The rules are originally for Dataset encoder. As far as I know, no maintainstream DBMS is using this policy by default.

Currently, the V1 data source uses "Legacy" policy by default, while V2 uses "Strict". This proposal is to use "ANSI" policy by default for both V1 and V2 in Spark 3.0.

There was also a DISCUSS thread "Follow ANSI SQL on table insertion" in the dev mailing list.

This vote is open until next Thurs (Sept. 12nd).

[ ] +1: Accept the proposal
[ ] +0
[ ] -1: I don't think this is a good idea because ...
Thank you!
Gengliang



--
Ryan Blue
Software Engineer
Netflix
Reply | Threaded
Open this post in threaded view
|

Re: [VOTE][SPARK-28885] Follow ANSI store assignment rules in table insertion by default

Felix Cheung
I’d prefer strict mode and fail fast (analysis check)

Also I like what Alastair suggested about standard clarification.

I think we can re-visit this proposal and restart the vote


From: Ryan Blue <[hidden email]>
Sent: Friday, September 6, 2019 5:28 PM
To: Alastair Green
Cc: Reynold Xin; Wenchen Fan; Spark dev list; Gengliang Wang
Subject: Re: [VOTE][SPARK-28885] Follow ANSI store assignment rules in table insertion by default
 

We discussed this thread quite a bit in the DSv2 sync up and Russell brought up a really good point about this.

The ANSI rule used here specifies how to store a specific value, V, so this is a runtime rule — an earlier case covers when V is NULL, so it is definitely referring to a specific value. The rule requires that if the type doesn’t match or if the value cannot be truncated, an exception is thrown for “numeric value out of range”.

That runtime error guarantees that even though the cast is introduced at analysis time, unexpected NULL values aren’t inserted into a table in place of data values that are out of range. Unexpected NULL values are the problem that was concerning to many of us in the discussion thread, but it turns out that real ANSI behavior doesn’t have the problem. (In the sync, we validated this by checking Postgres and MySQL behavior, too.)

In Spark, the runtime check is a separate configuration property from this one, but in order to actually implement ANSI semantics, both need to be set. So I think it makes sense tochange both defaults to be ANSI. The analysis check alone does not implement the ANSI standard.

In the sync, we also agreed that it makes sense to be able to turn off the runtime check in order to avoid job failures. Another, safer way to avoid job failures is to require an explicit cast, i.e., strict mode.

I think that we should amend this proposal to change the default for both the runtime check and the analysis check to ANSI.

As this stands now, I vote -1. But I would support this if the vote were to set both runtime and analysis checks to ANSI mode.

rb


On Fri, Sep 6, 2019 at 3:12 AM Alastair Green <[hidden email]> wrote:
Makes sense.

While the ISO SQL standard automatically becomes an American national  (ANSI) standard, changes are only made to the International (ISO/IEC) Standard, which is the authoritative specification. 

These rules are specified in SQL/Foundation (ISO/IEC SQL Part 2), section 9.2. 

Could we rename the proposed default to “ISO/IEC (ANSI)”? 

— Alastair

On Thu, Sep 5, 2019 at 17:17, Reynold Xin <[hidden email]> wrote:
Having three modes is a lot. Why not just use ansi mode as default, and legacy for backward compatibility? Then over time there's only the ANSI mode, which is standard compliant and easy to understand. We also don't need to invent a standard just for Spark.


On Thu, Sep 05, 2019 at 12:27 AM, Wenchen Fan <[hidden email]> wrote:
+1

To be honest I don't like the legacy policy. It's too loose and easy for users to make mistakes, especially when Spark returns null if a function hit errors like overflow.

The strict policy is not good either. It's too strict and stops valid use cases like writing timestamp values to a date type column. Users do expect truncation to happen without adding cast manually in this case. It's also weird to use a spark specific policy that no other database is using.

The ANSI policy is better. It stops invalid use cases like writing string values to an int type column, while keeping valid use cases like timestamp -> date.

I think it's no doubt that we should use ANSI policy instead of legacy policy for v1 tables. Except for backward compatibility, ANSI policy is literally better than the legacy policy.

The v2 table is arguable here. Although the ANSI policy is better than strict policy to me, this is just the store assignment policy, which only partially controls the table insertion behavior. With Spark's "return null on error" behavior, the table insertion is more likely to insert invalid null values with the ANSI policy compared to the strict policy.

I think we should use ANSI policy by default for both v1 and v2 tables, because
1. End-users don't care how the table is implemented. Spark should provide consistent table insertion behavior between v1 and v2 tables.
2. Data Source V2 is unstable in Spark 2.x so there is no backward compatibility issue. That said, the baseline to judge which policy is better should be the table insertion behavior in Spark 2.x, which is the legacy policy + "return null on error". ANSI policy is better than the baseline.
3. We expect more and more uses to migrate their data sources to the V2 API. The strict policy can be a stopper as it's a too big breaking change, which may break many existing queries.

Thanks,
Wenchen


On Wed, Sep 4, 2019 at 1:59 PM Gengliang Wang <[hidden email]> wrote:
Hi everyone,

I'd like to call for a vote on SPARK-28885 "Follow ANSI store assignment rules in table insertion by default".
When inserting a value into a column with the different data type, Spark performs type coercion. Currently, we support 3 policies for the type coercion rules: ANSI, legacy and strict, which can be set via the option "spark.sql.storeAssignmentPolicy":
1. ANSI: Spark performs the type coercion as per ANSI SQL. In practice, the behavior is mostly the same as PostgreSQL. It disallows certain unreasonable type conversions such as converting `string` to `int` and `double` to `boolean`.
2. Legacy: Spark allows the type coercion as long as it is a valid `Cast`, which is very loose. E.g., converting either `string` to `int` or `double` to `boolean` is allowed. It is the current behavior in Spark 2.x for compatibility with Hive.
3. Strict: Spark doesn't allow any possible precision loss or data truncation in type coercion, e.g., converting either `double` to `int` or `decimal` to `double` is allowed. The rules are originally for Dataset encoder. As far as I know, no maintainstream DBMS is using this policy by default.

Currently, the V1 data source uses "Legacy" policy by default, while V2 uses "Strict". This proposal is to use "ANSI" policy by default for both V1 and V2 in Spark 3.0.

There was also a DISCUSS thread "Follow ANSI SQL on table insertion" in the dev mailing list.

This vote is open until next Thurs (Sept. 12nd).

[ ] +1: Accept the proposal
[ ] +0
[ ] -1: I don't think this is a good idea because ...
Thank you!
Gengliang



--
Ryan Blue
Software Engineer
Netflix
Reply | Threaded
Open this post in threaded view
|

Re: [VOTE][SPARK-28885] Follow ANSI store assignment rules in table insertion by default

cloud0fan
I think it's too risky to enable the "runtime exception" mode by default in the next release. We don't even have a spec to describe when Spark would throw runtime exceptions. Currently the "runtime exception" mode works for overflow but I believe there are more places need to be considered (e.g. divide by zero).

However, Ryan has a good point that if we use the ANSI store assignment policy, we should make sure the table insertion behavior completely follows the SQL spec. After reading the related section in the SQL spec, the rule is to throw runtime exception for value out of range, which is the overflow check we already have in Spark. I think we should enable the overflow check during table insertion, when ANSI policy is picked. This should be done no matter which policy becomes the default eventually.

On Mon, Sep 9, 2019 at 8:00 AM Felix Cheung <[hidden email]> wrote:
I’d prefer strict mode and fail fast (analysis check)

Also I like what Alastair suggested about standard clarification.

I think we can re-visit this proposal and restart the vote


From: Ryan Blue <[hidden email]>
Sent: Friday, September 6, 2019 5:28 PM
To: Alastair Green
Cc: Reynold Xin; Wenchen Fan; Spark dev list; Gengliang Wang
Subject: Re: [VOTE][SPARK-28885] Follow ANSI store assignment rules in table insertion by default
 

We discussed this thread quite a bit in the DSv2 sync up and Russell brought up a really good point about this.

The ANSI rule used here specifies how to store a specific value, V, so this is a runtime rule — an earlier case covers when V is NULL, so it is definitely referring to a specific value. The rule requires that if the type doesn’t match or if the value cannot be truncated, an exception is thrown for “numeric value out of range”.

That runtime error guarantees that even though the cast is introduced at analysis time, unexpected NULL values aren’t inserted into a table in place of data values that are out of range. Unexpected NULL values are the problem that was concerning to many of us in the discussion thread, but it turns out that real ANSI behavior doesn’t have the problem. (In the sync, we validated this by checking Postgres and MySQL behavior, too.)

In Spark, the runtime check is a separate configuration property from this one, but in order to actually implement ANSI semantics, both need to be set. So I think it makes sense tochange both defaults to be ANSI. The analysis check alone does not implement the ANSI standard.

In the sync, we also agreed that it makes sense to be able to turn off the runtime check in order to avoid job failures. Another, safer way to avoid job failures is to require an explicit cast, i.e., strict mode.

I think that we should amend this proposal to change the default for both the runtime check and the analysis check to ANSI.

As this stands now, I vote -1. But I would support this if the vote were to set both runtime and analysis checks to ANSI mode.

rb


On Fri, Sep 6, 2019 at 3:12 AM Alastair Green <[hidden email]> wrote:
Makes sense.

While the ISO SQL standard automatically becomes an American national  (ANSI) standard, changes are only made to the International (ISO/IEC) Standard, which is the authoritative specification. 

These rules are specified in SQL/Foundation (ISO/IEC SQL Part 2), section 9.2. 

Could we rename the proposed default to “ISO/IEC (ANSI)”? 

— Alastair

On Thu, Sep 5, 2019 at 17:17, Reynold Xin <[hidden email]> wrote:
Having three modes is a lot. Why not just use ansi mode as default, and legacy for backward compatibility? Then over time there's only the ANSI mode, which is standard compliant and easy to understand. We also don't need to invent a standard just for Spark.


On Thu, Sep 05, 2019 at 12:27 AM, Wenchen Fan <[hidden email]> wrote:
+1

To be honest I don't like the legacy policy. It's too loose and easy for users to make mistakes, especially when Spark returns null if a function hit errors like overflow.

The strict policy is not good either. It's too strict and stops valid use cases like writing timestamp values to a date type column. Users do expect truncation to happen without adding cast manually in this case. It's also weird to use a spark specific policy that no other database is using.

The ANSI policy is better. It stops invalid use cases like writing string values to an int type column, while keeping valid use cases like timestamp -> date.

I think it's no doubt that we should use ANSI policy instead of legacy policy for v1 tables. Except for backward compatibility, ANSI policy is literally better than the legacy policy.

The v2 table is arguable here. Although the ANSI policy is better than strict policy to me, this is just the store assignment policy, which only partially controls the table insertion behavior. With Spark's "return null on error" behavior, the table insertion is more likely to insert invalid null values with the ANSI policy compared to the strict policy.

I think we should use ANSI policy by default for both v1 and v2 tables, because
1. End-users don't care how the table is implemented. Spark should provide consistent table insertion behavior between v1 and v2 tables.
2. Data Source V2 is unstable in Spark 2.x so there is no backward compatibility issue. That said, the baseline to judge which policy is better should be the table insertion behavior in Spark 2.x, which is the legacy policy + "return null on error". ANSI policy is better than the baseline.
3. We expect more and more uses to migrate their data sources to the V2 API. The strict policy can be a stopper as it's a too big breaking change, which may break many existing queries.

Thanks,
Wenchen


On Wed, Sep 4, 2019 at 1:59 PM Gengliang Wang <[hidden email]> wrote:
Hi everyone,

I'd like to call for a vote on SPARK-28885 "Follow ANSI store assignment rules in table insertion by default".
When inserting a value into a column with the different data type, Spark performs type coercion. Currently, we support 3 policies for the type coercion rules: ANSI, legacy and strict, which can be set via the option "spark.sql.storeAssignmentPolicy":
1. ANSI: Spark performs the type coercion as per ANSI SQL. In practice, the behavior is mostly the same as PostgreSQL. It disallows certain unreasonable type conversions such as converting `string` to `int` and `double` to `boolean`.
2. Legacy: Spark allows the type coercion as long as it is a valid `Cast`, which is very loose. E.g., converting either `string` to `int` or `double` to `boolean` is allowed. It is the current behavior in Spark 2.x for compatibility with Hive.
3. Strict: Spark doesn't allow any possible precision loss or data truncation in type coercion, e.g., converting either `double` to `int` or `decimal` to `double` is allowed. The rules are originally for Dataset encoder. As far as I know, no maintainstream DBMS is using this policy by default.

Currently, the V1 data source uses "Legacy" policy by default, while V2 uses "Strict". This proposal is to use "ANSI" policy by default for both V1 and V2 in Spark 3.0.

There was also a DISCUSS thread "Follow ANSI SQL on table insertion" in the dev mailing list.

This vote is open until next Thurs (Sept. 12nd).

[ ] +1: Accept the proposal
[ ] +0
[ ] -1: I don't think this is a good idea because ...
Thank you!
Gengliang



--
Ryan Blue
Software Engineer
Netflix
Reply | Threaded
Open this post in threaded view
|

Re: [VOTE][SPARK-28885] Follow ANSI store assignment rules in table insertion by default

Gengliang Wang
Thanks for the great suggestions from Ryan, Russell, and Wenchen.
As there are -1 from Ryan and Felix, this vote doesn't pass. 

As per the SQL standard, data rounding or truncation is allowed on assigning value to numeric/datetime type. So, I think we can discuss whether data rounding/truncation should be allowed in strict mode, as Spark doesn't produce invalid null values with data rounding/truncation.
For example, the conversion between `Date` and `Timestamp` should be allowed as per the SQL standard. For another example, converting `Decimal` to `Double` should be allowed as well, since in Spark SQL the max precision of Decimal type is 38, while the range of `Double` is [-1.7976931348623157E308, 1.7976931348623157E308]. Converting `Decimal` to `Double` won't cause an overflow in Spark.

After refining the ANSI mode and strict modes, we can vote for the default table insertion behavior for both V1 and V2.



On Thu, Sep 12, 2019 at 2:09 PM Wenchen Fan <[hidden email]> wrote:
I think it's too risky to enable the "runtime exception" mode by default in the next release. We don't even have a spec to describe when Spark would throw runtime exceptions. Currently the "runtime exception" mode works for overflow but I believe there are more places need to be considered (e.g. divide by zero).

However, Ryan has a good point that if we use the ANSI store assignment policy, we should make sure the table insertion behavior completely follows the SQL spec. After reading the related section in the SQL spec, the rule is to throw runtime exception for value out of range, which is the overflow check we already have in Spark. I think we should enable the overflow check during table insertion, when ANSI policy is picked. This should be done no matter which policy becomes the default eventually.

On Mon, Sep 9, 2019 at 8:00 AM Felix Cheung <[hidden email]> wrote:
I’d prefer strict mode and fail fast (analysis check)

Also I like what Alastair suggested about standard clarification.

I think we can re-visit this proposal and restart the vote


From: Ryan Blue <[hidden email]>
Sent: Friday, September 6, 2019 5:28 PM
To: Alastair Green
Cc: Reynold Xin; Wenchen Fan; Spark dev list; Gengliang Wang
Subject: Re: [VOTE][SPARK-28885] Follow ANSI store assignment rules in table insertion by default
 

We discussed this thread quite a bit in the DSv2 sync up and Russell brought up a really good point about this.

The ANSI rule used here specifies how to store a specific value, V, so this is a runtime rule — an earlier case covers when V is NULL, so it is definitely referring to a specific value. The rule requires that if the type doesn’t match or if the value cannot be truncated, an exception is thrown for “numeric value out of range”.

That runtime error guarantees that even though the cast is introduced at analysis time, unexpected NULL values aren’t inserted into a table in place of data values that are out of range. Unexpected NULL values are the problem that was concerning to many of us in the discussion thread, but it turns out that real ANSI behavior doesn’t have the problem. (In the sync, we validated this by checking Postgres and MySQL behavior, too.)

In Spark, the runtime check is a separate configuration property from this one, but in order to actually implement ANSI semantics, both need to be set. So I think it makes sense tochange both defaults to be ANSI. The analysis check alone does not implement the ANSI standard.

In the sync, we also agreed that it makes sense to be able to turn off the runtime check in order to avoid job failures. Another, safer way to avoid job failures is to require an explicit cast, i.e., strict mode.

I think that we should amend this proposal to change the default for both the runtime check and the analysis check to ANSI.

As this stands now, I vote -1. But I would support this if the vote were to set both runtime and analysis checks to ANSI mode.

rb


On Fri, Sep 6, 2019 at 3:12 AM Alastair Green <[hidden email]> wrote:
Makes sense.

While the ISO SQL standard automatically becomes an American national  (ANSI) standard, changes are only made to the International (ISO/IEC) Standard, which is the authoritative specification. 

These rules are specified in SQL/Foundation (ISO/IEC SQL Part 2), section 9.2. 

Could we rename the proposed default to “ISO/IEC (ANSI)”? 

— Alastair

On Thu, Sep 5, 2019 at 17:17, Reynold Xin <[hidden email]> wrote:
Having three modes is a lot. Why not just use ansi mode as default, and legacy for backward compatibility? Then over time there's only the ANSI mode, which is standard compliant and easy to understand. We also don't need to invent a standard just for Spark.


On Thu, Sep 05, 2019 at 12:27 AM, Wenchen Fan <[hidden email]> wrote:
+1

To be honest I don't like the legacy policy. It's too loose and easy for users to make mistakes, especially when Spark returns null if a function hit errors like overflow.

The strict policy is not good either. It's too strict and stops valid use cases like writing timestamp values to a date type column. Users do expect truncation to happen without adding cast manually in this case. It's also weird to use a spark specific policy that no other database is using.

The ANSI policy is better. It stops invalid use cases like writing string values to an int type column, while keeping valid use cases like timestamp -> date.

I think it's no doubt that we should use ANSI policy instead of legacy policy for v1 tables. Except for backward compatibility, ANSI policy is literally better than the legacy policy.

The v2 table is arguable here. Although the ANSI policy is better than strict policy to me, this is just the store assignment policy, which only partially controls the table insertion behavior. With Spark's "return null on error" behavior, the table insertion is more likely to insert invalid null values with the ANSI policy compared to the strict policy.

I think we should use ANSI policy by default for both v1 and v2 tables, because
1. End-users don't care how the table is implemented. Spark should provide consistent table insertion behavior between v1 and v2 tables.
2. Data Source V2 is unstable in Spark 2.x so there is no backward compatibility issue. That said, the baseline to judge which policy is better should be the table insertion behavior in Spark 2.x, which is the legacy policy + "return null on error". ANSI policy is better than the baseline.
3. We expect more and more uses to migrate their data sources to the V2 API. The strict policy can be a stopper as it's a too big breaking change, which may break many existing queries.

Thanks,
Wenchen


On Wed, Sep 4, 2019 at 1:59 PM Gengliang Wang <[hidden email]> wrote:
Hi everyone,

I'd like to call for a vote on SPARK-28885 "Follow ANSI store assignment rules in table insertion by default".
When inserting a value into a column with the different data type, Spark performs type coercion. Currently, we support 3 policies for the type coercion rules: ANSI, legacy and strict, which can be set via the option "spark.sql.storeAssignmentPolicy":
1. ANSI: Spark performs the type coercion as per ANSI SQL. In practice, the behavior is mostly the same as PostgreSQL. It disallows certain unreasonable type conversions such as converting `string` to `int` and `double` to `boolean`.
2. Legacy: Spark allows the type coercion as long as it is a valid `Cast`, which is very loose. E.g., converting either `string` to `int` or `double` to `boolean` is allowed. It is the current behavior in Spark 2.x for compatibility with Hive.
3. Strict: Spark doesn't allow any possible precision loss or data truncation in type coercion, e.g., converting either `double` to `int` or `decimal` to `double` is allowed. The rules are originally for Dataset encoder. As far as I know, no maintainstream DBMS is using this policy by default.

Currently, the V1 data source uses "Legacy" policy by default, while V2 uses "Strict". This proposal is to use "ANSI" policy by default for both V1 and V2 in Spark 3.0.

There was also a DISCUSS thread "Follow ANSI SQL on table insertion" in the dev mailing list.

This vote is open until next Thurs (Sept. 12nd).

[ ] +1: Accept the proposal
[ ] +0
[ ] -1: I don't think this is a good idea because ...
Thank you!
Gengliang



--
Ryan Blue
Software Engineer
Netflix