Thread: Plans for 2.8
Hello, next week I will have some time and maybe could end all the work to release psycopg 2.8. I am aware there are several features which have been awaiting for some time. The state of the work can be seen in [1]. [1] https://github.com/psycopg/psycopg2/issues?q=milestone%3A"psycopg+2.8" Psycopg 2.8 will only support Python 2.7 and 3.4+. The codebase has been heavily hammered by Jon Dufresne who has killed a lot of Py2isms and the whole use of 2to3 is now replaced by a minimal compatibility module. So Python is now as modern as C in supporting both 2 and 3 with a single codebase :P Other deprecated and unused objects have also been dropped, see the news file [2]. Fog, if you can take a look at examples/sandbox and delete what's no more required there (#645), that would be great :) [2] https://github.com/psycopg/psycopg2/blob/master/NEWS The feature I'm the most excited about (and worried about its reception) is to raise a different exception for every postgres error message (see #682) . For instance `SELECT * FROM wrong_name` will raise `UndefinedTable` rather than `ProgrammingError`. Currently handling a specific exception requires catching a broader class and looking at the pgcode: try: cur.execute("lock table %s in access exclusive mode nowait" % name) except psycopg2.OperationalError as e: if e.pgcode == psycopg2.errorcodes.LOCK_NOT_AVAILABLE: locked = True else: raise This can become a much more natural: try: cur.execute("lock table %s in access exclusive mode nowait" % name) except psycopg2.errors.LockNotAvailable: locked = True The error classes are generated automatically from Postgres source code and are subclasses of the previously existing ones, so existing code should be unaffected. I'd be happy to have input about the feature and suggestions before releasing it. A tiny improvement to SQL generation is already ready^W merged in #732: it will be possible to use `Identifier("schema", "name")` which would be rendered in dotted notation in the query. Currently `Identifier()` takes a single param so this extension is backward compatible and there is no need to introduce a new `Composable` type to represent dotted sequences of identifiers. There are requests to get extra info about the connection or the result (see #726, #661). They are reasonable and not too difficult to implement so I'd like to give them a go. However they are easy enough for someone to contribute if you feel? That would be very appreciated and would reduce the surface of the works to perform on my part. Another tiny feature would be to support IntEnum out-of-the-box (#591), which I've never used in Python. In the other thread these days we have discussed about introducing capsules: we can take a look to that too... Added #782. Thank you very much for any contribution, with ideas and even more with code :) -- Daniele
On 10/04/2018 02:38 PM, Daniele Varrazzo wrote: > A tiny improvement to SQL generation is already ready^W merged in > #732: it will be possible to use `Identifier("schema", "name")` which > would be rendered in dotted notation in the query. Currently > `Identifier()` takes a single param so this extension is backward > compatible and there is no need to introduce a new `Composable` type > to represent dotted sequences of identifiers. I understand that from a compatibility point of view everything works with the "schema", "name" order of arguments (you just switch on the number of arguments) but usually such approach causes infinite headaches when you remove or add the namespace from the call. `Identifier(name, schema=None)` is better, IMHO because makes explicit that the mandatory and first argument is always the identifier itself, while the schema is optional. federico -- Federico Di Gregorio federico.digregorio@dndg.it DNDG srl http://dndg.it One key. One input. One enter. All right. -- An american consultant (then the system crashed and took down the *entire* network)
On 04/10/18, Daniele Varrazzo (daniele.varrazzo@gmail.com) wrote: > The feature I'm the most excited about (and worried about its > reception) is to raise a different exception for every postgres error > message (see #682) . For instance `SELECT * FROM wrong_name` will > raise `UndefinedTable` rather than `ProgrammingError`. Currently > handling a specific exception requires catching a broader class and > looking at the pgcode: > > try: > cur.execute("lock table %s in access exclusive mode nowait" % name) > except psycopg2.OperationalError as e: > if e.pgcode == psycopg2.errorcodes.LOCK_NOT_AVAILABLE: > locked = True > else: > raise > > This can become a much more natural: > > try: > cur.execute("lock table %s in access exclusive mode nowait" % name) > except psycopg2.errors.LockNotAvailable: > locked = True > > The error classes are generated automatically from Postgres source > code and are subclasses of the previously existing ones, so existing > code should be unaffected. I'd be happy to have input about the > feature and suggestions before releasing it. Hi Daniele The greater depth of exception reporting looks great to me, particularly if they are subclasses of the existing ones. Regards Rory
On Thu, Oct 4, 2018 at 2:27 PM Federico Di Gregorio <fog@dndg.it> wrote: > > On 10/04/2018 02:38 PM, Daniele Varrazzo wrote: > > A tiny improvement to SQL generation is already ready^W merged in > > #732: it will be possible to use `Identifier("schema", "name")` which > > would be rendered in dotted notation in the query. Currently > > `Identifier()` takes a single param so this extension is backward > > compatible and there is no need to introduce a new `Composable` type > > to represent dotted sequences of identifiers. > > I understand that from a compatibility point of view everything works > with the "schema", "name" order of arguments (you just switch on the > number of arguments) but usually such approach causes infinite headaches > when you remove or add the namespace from the call. > > `Identifier(name, schema=None)` is better, IMHO because makes explicit > that the mandatory and first argument is always the identifier itself, > while the schema is optional. "schema", "table" is only an example: it could be "table"."field", even "schema"."table"."field", or "extension"."setting"... The object only wants to represent a dotted sequence of identifiers, at lexical level, nothing with semantics attached such as "an optionally schema-qualified table name" or "a field name". If the object was `Table()` or `Field()` rather than `Identifier()` I'd totally agree with you. -- Daniele
On 10/04/2018 04:05 PM, Daniele Varrazzo wrote: > On Thu, Oct 4, 2018 at 2:27 PM Federico Di Gregorio<fog@dndg.it> wrote: >> On 10/04/2018 02:38 PM, Daniele Varrazzo wrote: >>> A tiny improvement to SQL generation is already ready^W merged in >>> #732: it will be possible to use `Identifier("schema", "name")` which >>> would be rendered in dotted notation in the query. Currently >>> `Identifier()` takes a single param so this extension is backward >>> compatible and there is no need to introduce a new `Composable` type >>> to represent dotted sequences of identifiers. >> I understand that from a compatibility point of view everything works >> with the "schema", "name" order of arguments (you just switch on the >> number of arguments) but usually such approach causes infinite headaches >> when you remove or add the namespace from the call. >> >> `Identifier(name, schema=None)` is better, IMHO because makes explicit >> that the mandatory and first argument is always the identifier itself, >> while the schema is optional. > "schema", "table" is only an example: it could be "table"."field", > even "schema"."table"."field", or "extension"."setting"... The object > only wants to represent a dotted sequence of identifiers, at lexical > level, nothing with semantics attached such as "an optionally > schema-qualified table name" or "a field name". If the object was > `Table()` or `Field()` rather than `Identifier()` I'd totally agree > with you. Sorry, I misread your example. Obviously you're right. federico p.s. yep, I'll remove all the old cruft from sandbox. -- Federico Di Gregorio federico.digregorio@dndg.it DNDG srl http://dndg.it The number of the beast: vi vi vi. -- Delexa Jones
On Thu, Oct 4, 2018 at 8:38 AM Daniele Varrazzo <daniele.varrazzo@gmail.com> wrote: > > Hello, > > next week I will have some time and maybe could end all the work to > release psycopg 2.8. I am aware there are several features which have > been awaiting for some time. The state of the work can be seen in [1]. > > [1] https://github.com/psycopg/psycopg2/issues?q=milestone%3A"psycopg+2.8" > > Psycopg 2.8 will only support Python 2.7 and 3.4+. The codebase has > been heavily hammered by Jon Dufresne who has killed a lot of Py2isms > and the whole use of 2to3 is now replaced by a minimal compatibility > module. So Python is now as modern as C in supporting both 2 and 3 > with a single codebase :P Other deprecated and unused objects have > also been dropped, see the news file [2]. Fog, if you can take a look > at examples/sandbox and delete what's no more required there (#645), > that would be great :) > > [2] https://github.com/psycopg/psycopg2/blob/master/NEWS > > The feature I'm the most excited about (and worried about its > reception) is to raise a different exception for every postgres error > message (see #682) . For instance `SELECT * FROM wrong_name` will > raise `UndefinedTable` rather than `ProgrammingError`. Currently > handling a specific exception requires catching a broader class and > looking at the pgcode: > > try: > cur.execute("lock table %s in access exclusive mode nowait" % name) > except psycopg2.OperationalError as e: > if e.pgcode == psycopg2.errorcodes.LOCK_NOT_AVAILABLE: > locked = True > else: > raise > > This can become a much more natural: > > try: > cur.execute("lock table %s in access exclusive mode nowait" % name) > except psycopg2.errors.LockNotAvailable: > locked = True > > The error classes are generated automatically from Postgres source > code and are subclasses of the previously existing ones, so existing > code should be unaffected. I'd be happy to have input about the > feature and suggestions before releasing it. I can't provide any suggestions, as the feature is very reasonable and useful. But I will lament that pep-249 has nothing about this, which means from a driver-agnostic point of view, the situation is pretty much unchanged. Here's code I wrote for Openstack to try to apply more specificity to database errors, basically a library of regexes: https://github.com/openstack/oslo.db/blob/master/oslo_db/sqlalchemy/exc_filters.py#L55 one thing that would be helpful would be if your fine-grained exception classes included more context about the failure. Like UndefinedTable would include the table name as an individual datamember e.g. exception.table_name, an error about a foreign key constraint would include the constraint name e.g. exception.constraint_name, things like that. You can see in my oslo.db library above we are also pulling out other elements from the error message to provide more context. > > A tiny improvement to SQL generation is already ready^W merged in > #732: it will be possible to use `Identifier("schema", "name")` which > would be rendered in dotted notation in the query. Currently > `Identifier()` takes a single param so this extension is backward > compatible and there is no need to introduce a new `Composable` type > to represent dotted sequences of identifiers. > > There are requests to get extra info about the connection or the > result (see #726, #661). They are reasonable and not too difficult to > implement so I'd like to give them a go. However they are easy enough > for someone to contribute if you feel? That would be very appreciated > and would reduce the surface of the works to perform on my part. > Another tiny feature would be to support IntEnum out-of-the-box > (#591), which I've never used in Python. > > In the other thread these days we have discussed about introducing > capsules: we can take a look to that too... Added #782. > > Thank you very much for any contribution, with ideas and even more with code :) > > -- Daniele >
On Thu, Oct 4, 2018 at 3:18 PM Mike Bayer <mike_mp@zzzcomputing.com> wrote: > I can't provide any suggestions, as the feature is very reasonable and > useful. But I will lament that pep-249 has nothing about this, which > means from a driver-agnostic point of view, the situation is pretty > much unchanged. Here's code I wrote for Openstack to try to apply > more specificity to database errors, basically a library of regexes: > https://github.com/openstack/oslo.db/blob/master/oslo_db/sqlalchemy/exc_filters.py#L55 Uhm, they also have the problem of not working if the message is localised... :\ With postgres it woud have been more robust to look at the extension `pgcode` but of course that's not portable either. Is there anything in common to all the databases which might be exposed in an uniform way by the drivers? e.g. the pgcode is actually something more standard than just postgres: "sqlstate" (or, SQLSTATE, because '70s) is supposed to be a standard. If you know that many/all the database emit a sqlstate you may suggest the dbsig to bless an exception attribute - e.g. `Exception.sqlstate` - to report it. Of course if postgres says "40P01" and IBM DB2 says "0911N" to report a deadlock, that's way out of what we can control... > one thing that would be helpful would be if your fine-grained > exception classes included more context about the failure. Like > UndefinedTable would include the table name as an individual > datamember e.g. exception.table_name, an error about a foreign key > constraint would include the constraint name e.g. > exception.constraint_name, things like that. You can see in my > oslo.db library above we are also pulling out other elements from the > error message to provide more context. We do already: more details about the exception, if made available by the database, are made available by the exception `diag` attribute: see <http://initd.org/psycopg/docs/extensions.html#psycopg2.extensions.Diagnostics>. -- Daniele
On Thu, Oct 04, 2018 at 01:38:03PM +0100, Daniele Varrazzo wrote: > [2] https://github.com/psycopg/psycopg2/blob/master/NEWS > > The feature I'm the most excited about (and worried about its > reception) is to raise a different exception for every postgres error > message (see #682). You needn't as far as I'm concerned because as long as this > The error classes are [...] subclasses of the previously existing ones is true there should not be a problem. Will the new classes still have .pgcode set to their respective values ? Thanks, Karsten -- GPG 40BE 5B0E C98E 1713 AFA6 5BC0 3BEA AC80 7D4F C89B