[WIP][RFC]sequences support

Question

[WIP][RFC]sequences support

Closed this issue 7 years ago · 4 comments

No BC breaks.

Thought process history

While reviewing #172, helping with #162, and fresh out of a heavy database driven enterprise platform contract, noticed how deficient current sequence support is plus misleading function signatures. This is an important feature because, better than auto_increment in PostgreSQL, the only way to increment Oracle, and now Microsoft brought it in due to popular demand in SQL Server 2016 so I feel it needs detailed look.

Current situation is that:

TableGateway can only deal with a single sequence instance primarily for primary key usage, despite accepting unrestricted number of features.
As pointed out by @andrey-mokhov only supports explicit sequence naming. In some engines like Oracle, MSSQL2016 this is fine, but PostgreSQL supports auto-generated sequence names with SERIAL which is a problem.
Discovered at #191 that RowGateway is also incapable of handling sequences. Limitation is in fact that RowGateway does not link to any table gateway. Easy to add as an optional dependency by making constructor support TableGateway along with string and TableIdentifier OR add addition of features like other db classes, but hard in that INSERT operation on new record cannot be engine augnostic. May need to explicitly check for existence of that feature even if it breaks good OOP practice (or better suggestion).

What I have seen in the field:

Sequences are not always used for primary key generation. I experienced this myself, is documented at https://msdn.microsoft.com/en-us/library/ff878058.aspx and requested in our Hibernate neighbours http://stackoverflow.com/questions/277630/hibernate-jpa-sequence-non-id
Multiple sequences for a table, one for primary key, others for auto-incrementing based on some business rule, are not uncommon. Think invoice number generator across multiple sales departments, worker queue management, etc.
Mix of implicit sequence names (usually for PK) and explicitly named in schema (usually for business rules)

Submitted PR #187 where I got around multiple features of same type problem. Intended usage would be (documented of course when settled one one)

$tableGateway = new TableGateway('table', $adapter, [new Sequence('pk_column', 'seq_1'), new Sequence('seq_2')]);
$tableGateway->nextSequenceId(); // no parameter aimed at 'seq_1' implying primary key
$tableGateway->nextSequenceId('seq_2'); // parameter specifying which one of the participating sequences are interested in

Why not always let developers manually handle Sequence objects without attaching them to table gateway? In large schemas, relationships set by architect can be difficult to keep in mind and I caused horrible bugs not being aware of what sequence participates in what business logic transaction. Using tablegateway to group them together helps a lot.

Now undecided on how to deal with preInsert/postInsert and dependence on gateway, how would I optionally mark one of several sequences that is used in TableGateway as responsible for PK? How would lastInsertValue be managed for multiple autoincremented columns?

Answer 1 · 2017-02-05T15:41:35.000Z

Trying to consolidate conflicting sequence PRs reported over time in Zend DB. Stuck at deciding how to do platform specific code. Would like opinion of significantly more knowledgeable than I am individuals @froschdesign @Ocramius @weierophinney

Original code had switch/case statement per platform. Thats what bug reports focused on so I followed same pattern: https://github.com/zendframework/zend-db/pull/187/files#diff-85fe938f5a6f87443ad2578472eed9e0R166

But I just noticed there is duplication of that lastSequenceId() query in Connection class. Seems like a great benefit because do not need nasty switch($platform), and if anyone relied on similarly broken code in Connection too, all is fixed in one place.

https://github.com/zendframework/zend-db/blob/master/src/Adapter/Driver/Pgsql/Connection.php#L267

https://github.com/zendframework/zend-db/blob/master/src/Adapter/Driver/Pdo/Connection.php#L419

There are some concerns, however. What about nextSequenceId()? Connection interface does not have signature for it the way getLastGeneratedValue() is and rightfully so, I believe, should not be added. I could add nextSequenceId() to connection classes without adding to interface and assume dynamic typing will handle it, but that looks dirty. But code duplication looks dirty too, and so does inconsistency of doing switch/case for one, and call to adapter for the other method.

Another possible downside of using adapter connection, is getLastGeneratedValue() does not share same syntax for SqlServer 2016 where its one command for IDENTITY column but a different one for SEQUENCE object/column. (not a problem for postgresql or oracle up to 12c because sequence is the only way to do it)

Finally, how do I go about deducing implicit sequence names which too is a platform specific action? My getSequenceName() is awefully postgresql specific. I am not sure if decorator is possible to do for TableGateway features.

Answer 2 · 2017-02-05T15:45:33.000Z

oops, link to wrong branch. Corrected: https://github.com/zendframework/zend-db/pull/187/files#diff-85fe938f5a6f87443ad2578472eed9e0R166

Answer 3 · 2017-02-23T13:45:55.000Z

@ezimuel while you are triaging issues, do you have a minute to give a hint about the last question, please?

Answer 4 · 2017-12-07T16:45:49.000Z

Closing this in favor of keeping TODO and discussion in one place at linked PR.