COPY (query) TO file

Mark Woodward wrote:

Tom had posted a question about file compression with copy. I thought
about it, and I want to through this out and see if anyone things it is a
good idea.

Currently, the COPY command only copies a table, what if it could operate
with a query, as:

COPY (select * from mytable where foo='bar') as BAR TO stdout

I have no idea if it is doable, but I can see uses for replication

I doubt it be really usefull (apart from maybe saving some work
coding a client app) but did you actually test it with

create table as select ...; followed by a copy of that table
if it really is faster then just the usual select & fetch?

Regards
Tino

Mark Woodward wrote:

Tom had posted a question about file compression with copy. I thought
about it, and I want to through this out and see if anyone things it is a
good idea.

Currently, the COPY command only copies a table, what if it could operate
with a query, as:

COPY (select * from mytable where foo='bar') as BAR TO stdout

Isn't this already being worked on? The TODO list says:

Allow COPY to output from views
Another idea would be to allow actual SELECT statements in a COPY.

Personally I strongly favor the second option as being more flexible
than the first.

cheers

andrew

Mark Woodward wrote:

Tom had posted a question about file compression with copy. I thought
about it, and I want to through this out and see if anyone things it is
a
good idea.

Currently, the COPY command only copies a table, what if it could
operate
with a query, as:

COPY (select * from mytable where foo='bar') as BAR TO stdout

I have no idea if it is doable, but I can see uses for replication

I doubt it be really usefull (apart from maybe saving some work
coding a client app) but did you actually test it with

create table as select ...; followed by a copy of that table
if it really is faster then just the usual select & fetch?

Why "create table?"

The idea is that you would have one or more redundent databases and use
the COPY TO/FROM to keep them up to date.

Mark Woodward wrote:

Mark Woodward wrote:

Tom had posted a question about file compression with copy. I thought
about it, and I want to through this out and see if anyone things it is
a
good idea.

Currently, the COPY command only copies a table, what if it could
operate
with a query, as:

COPY (select * from mytable where foo='bar') as BAR TO stdout

I have no idea if it is doable, but I can see uses for replication

I doubt it be really usefull (apart from maybe saving some work
coding a client app) but did you actually test it with

create table as select ...; followed by a copy of that table
if it really is faster then just the usual select & fetch?

Why "create table?"

Just to simulate and time the proposal.
SELECT ... already works over the network and if COPY from a
select (which would basically work like yet another wire
protocol) isnt significantly faster, why bother?

The idea is that you would have one or more redundent databases and use
the COPY TO/FROM to keep them up to date.

Well, if you have databases you would have regular tables - and
can use copy as it is now :-)

Regards
Tino

Mark Woodward wrote:

Mark Woodward wrote:

Tom had posted a question about file compression with copy. I thought
about it, and I want to through this out and see if anyone things it
is
a
good idea.

Currently, the COPY command only copies a table, what if it could
operate
with a query, as:

COPY (select * from mytable where foo='bar') as BAR TO stdout

I have no idea if it is doable, but I can see uses for replication

I doubt it be really usefull (apart from maybe saving some work
coding a client app) but did you actually test it with

create table as select ...; followed by a copy of that table
if it really is faster then just the usual select & fetch?

Why "create table?"

Just to simulate and time the proposal.
SELECT ... already works over the network and if COPY from a
select (which would basically work like yet another wire
protocol) isnt significantly faster, why bother?

Because the format of COPY is a common transmiter/receiver for PostgreSQL,
like this:

pg_dump -t mytable | psql -h target -c "COPY mytable FROM STDIN"

With a more selective copy, you can use pretty much this mechanism to
limit a copy to a sumset of the records in a table.

The idea is that you would have one or more redundent databases and use
the COPY TO/FROM to keep them up to date.

Well, if you have databases you would have regular tables - and
can use copy as it is now :-)

But COPY copies all the records, not some of the records.

Andrew Dunstan wrote:

Mark Woodward wrote:

Tom had posted a question about file compression with copy. I thought
about it, and I want to through this out and see if anyone things it is a
good idea.

Currently, the COPY command only copies a table, what if it could operate
with a query, as:

COPY (select * from mytable where foo='bar') as BAR TO stdout

Isn't this already being worked on? The TODO list says:

Allow COPY to output from views
Another idea would be to allow actual SELECT statements in a COPY.

Personally I strongly favor the second option as being more flexible
than the first.

I second that - allowing arbitrary SELECT statements as a COPY source
seems much more powerful and flexible than just supporting COPY FROM VIEW.

Stefan

On Fri, 2006-06-02 at 09:56 -0400, Andrew Dunstan wrote:

Allow COPY to output from views

FYI, there is a patch for this floating around -- I believe it was
posted to -patches a few months back.

Another idea would be to allow actual SELECT statements in a COPY.

Personally I strongly favor the second option as being more flexible
than the first.

+1.

-Neil

Neil Conway wrote:

On Fri, 2006-06-02 at 09:56 -0400, Andrew Dunstan wrote:

Allow COPY to output from views

FYI, there is a patch for this floating around -- I believe it was
posted to -patches a few months back.

I have it. The pieces of it than I can use to implement the idea below,
I'll try to rescue. I don't think it's that much though.

Another idea would be to allow actual SELECT statements in a COPY.

Personally I strongly favor the second option as being more flexible
than the first.

+1.

Sounds like a plan.

--
Alvaro Herrera http://www.CommandPrompt.com/
PostgreSQL Replication, Consulting, Custom Development, 24x7 support

Mark Woodward wrote:
...

create table as select ...; followed by a copy of that table
if it really is faster then just the usual select & fetch?

Why "create table?"

Just to simulate and time the proposal.
SELECT ... already works over the network and if COPY from a
select (which would basically work like yet another wire
protocol) isnt significantly faster, why bother?

Because the format of COPY is a common transmiter/receiver for PostgreSQL,
like this:

pg_dump -t mytable | psql -h target -c "COPY mytable FROM STDIN"

With a more selective copy, you can use pretty much this mechanism to
limit a copy to a sumset of the records in a table.

Ok, but why not just implement this into pg_dump or psql?
Why bother the backend with that functionality?

For example if you copy numbers, int4 (4 bytes)
gets expanded to up to 10 bytes. Of course
can get the same wire load if you use to_char()
with regular select.

The idea is that you would have one or more redundent databases and use
the COPY TO/FROM to keep them up to date.

Well, if you have databases you would have regular tables - and
can use copy as it is now :-)

But COPY copies all the records, not some of the records.

yes, that would be the point in having them "up to date"
and not "partially maybe something up to date" ;)

COPY is fine for import of data, but for export
I think it should be implemented in the frontend.

Regards
Tino

Tino Wildenhain <tino@wildenhain.de> writes:

Ok, but why not just implement this into pg_dump or psql?
Why bother the backend with that functionality?

You're not seriously suggesting we reimplement evaluation of WHERE clauses
on the client side, are you?

regards, tom lane

Tom Lane wrote:

Tino Wildenhain <tino@wildenhain.de> writes:

Ok, but why not just implement this into pg_dump or psql?
Why bother the backend with that functionality?

You're not seriously suggesting we reimplement evaluation of WHERE clauses
on the client side, are you?

no, did I? But what is wrong with something like:

\COPY 'SELECT foo,bar,baz FROM footable WHERE baz=5 ORDER BY foo' TO
file|stdout

which would just run the query (in the backend of course) and
format the output just like copy would...

I mean, ok, when its in the backend, its fine too (beside the
data expansion if you dont implement compression...)
but I thougt implementing in the frontend would be easier...

Regards
Tino

Mark Woodward wrote:
...

create table as select ...; followed by a copy of that table
if it really is faster then just the usual select & fetch?

Why "create table?"

Just to simulate and time the proposal.
SELECT ... already works over the network and if COPY from a
select (which would basically work like yet another wire
protocol) isnt significantly faster, why bother?

Because the format of COPY is a common transmiter/receiver for
PostgreSQL,
like this:

pg_dump -t mytable | psql -h target -c "COPY mytable FROM STDIN"

With a more selective copy, you can use pretty much this mechanism to
limit a copy to a sumset of the records in a table.

Ok, but why not just implement this into pg_dump or psql?
Why bother the backend with that functionality?

Because "COPY" runs on the back-end, not the front end, and the front end
may not even be in the same city as the backend. When you issue a "COPY"
the file it reads or writes local to the backend. True, the examples I
gave may not show how that is important, but consider this:

psql -h remote masterdb -c "COPY (select * from mytable where ID <
xxlastxx) as mytable TO '/replicate_backup/mytable-060602.pgc'"

This runs completely in the background and can serve as a running backup.

Mark Woodward wrote:
...

pg_dump -t mytable | psql -h target -c "COPY mytable FROM STDIN"

With a more selective copy, you can use pretty much this mechanism to
limit a copy to a sumset of the records in a table.

Ok, but why not just implement this into pg_dump or psql?
Why bother the backend with that functionality?

Because "COPY" runs on the back-end, not the front end, and the front end
may not even be in the same city as the backend. When you issue a "COPY"
the file it reads or writes local to the backend. True, the examples I
gave may not show how that is important, but consider this:

We were talking about COPY to stdout :-) Copy to file is another
issue :-) Copy to (server fs) file has so many limitations I dont see
wide use for it. (Of course there are usecases)

psql -h remote masterdb -c "COPY (select * from mytable where ID <
xxlastxx) as mytable TO '/replicate_backup/mytable-060602.pgc'"

This runs completely in the background and can serve as a running backup.

And you are sure it would be much faster then a server local running
psql just dumping the result of a query?
(And you could more easy avoid raceconditions in contrast to several
remote clients trying to trigger your above backup )

But what do I know... I was just asking :-)

Regards
Tino

On Fri, Jun 02, 2006 at 09:56:07AM -0400, Andrew Dunstan wrote:

Mark Woodward wrote:

Tom had posted a question about file compression with copy. I thought
about it, and I want to through this out and see if anyone things it is a
good idea.

Currently, the COPY command only copies a table, what if it could operate
with a query, as:

COPY (select * from mytable where foo='bar') as BAR TO stdout

Isn't this already being worked on? The TODO list says:

Allow COPY to output from views
Another idea would be to allow actual SELECT statements in a COPY.

Personally I strongly favor the second option as being more flexible
than the first.

+1 :)

Cheers,
D
--
David Fetter <david@fetter.org> http://fetter.org/
phone: +1 415 235 3778 AIM: dfetter666
Skype: davidfetter

Remember to vote!

Tino Wildenhain <tino@wildenhain.de> writes:

Tom Lane wrote:

Tino Wildenhain <tino@wildenhain.de> writes:

Ok, but why not just implement this into pg_dump or psql?
Why bother the backend with that functionality?

You're not seriously suggesting we reimplement evaluation of WHERE clauses
on the client side, are you?

No, he's suggesting the client implement COPY formatting after fetching a
regular result set.

Of course this runs into the same problem other clients have dealing with
large result sets. libpq doesn't want to let the client deal with partial
results so you have to buffer up the entire result set in memory.

I was also vaguely pondering whether all the DDL commands could be generalized
to receive or send COPY formatted data for repeated execution. It would be
neat to be able to prepare an UPDATE with placeholders and stream data in COPY
format as parameters to the UPDATE to execute it thousands or millions of
times without any protocol overhead or network pipeline stalls.

--
greg

Mark Woodward wrote:
...

pg_dump -t mytable | psql -h target -c "COPY mytable FROM STDIN"

With a more selective copy, you can use pretty much this mechanism to
limit a copy to a sumset of the records in a table.

Ok, but why not just implement this into pg_dump or psql?
Why bother the backend with that functionality?

Because "COPY" runs on the back-end, not the front end, and the front
end
may not even be in the same city as the backend. When you issue a "COPY"
the file it reads or writes local to the backend. True, the examples I
gave may not show how that is important, but consider this:

We were talking about COPY to stdout :-) Copy to file is another
issue :-) Copy to (server fs) file has so many limitations I dont see
wide use for it. (Of course there are usecases)

"wide use for" is not always the same as "useful." Sometimes "useful" is
something not easily doable in other ways or completes a feature set.

psql -h remote masterdb -c "COPY (select * from mytable where ID <
xxlastxx) as mytable TO '/replicate_backup/mytable-060602.pgc'"

This runs completely in the background and can serve as a running
backup.

And you are sure it would be much faster then a server local running
psql just dumping the result of a query?

No I can't be sure of that at all, but .... The COPY command has a
specific use that is understood and an operation that is separate from the
normal query mechanism.

(And you could more easy avoid raceconditions in contrast to several
remote clients trying to trigger your above backup )

Again, the examples may not have been precise in presenting "why," the
focus was mostly "what" so it could be discussed. As a generic feature it
has many potential uses. Trying to debate and defend a specific use limits
the potential scope of the feature.

Why have COPY anyway? Why not just use "SELECT * FROM TABLE?"

Allow COPY to output from views
Another idea would be to allow actual SELECT statements in a COPY.

Personally I strongly favor the second option as being more flexible
than the first.

I second that - allowing arbitrary SELECT statements as a COPY source
seems much more powerful and flexible than just supporting COPY FROM VIEW.

Not to be a sour apple or anything but I don't see how any of this is
needed in the backend since we can easily use Psql to do it, or pretty
much any other tool.

Joshua D. Drake

Stefan

---------------------------(end of broadcast)---------------------------
TIP 6: explain analyze is your friend

--

=== The PostgreSQL Company: Command Prompt, Inc. ===
Sales/Support: +1.503.667.4564 || 24x7/Emergency: +1.800.492.2240
Providing the most comprehensive PostgreSQL solutions since 1997
http://www.commandprompt.com/

Allow COPY to output from views
Another idea would be to allow actual SELECT statements in a COPY.

Personally I strongly favor the second option as being more flexible
than the first.

I second that - allowing arbitrary SELECT statements as a COPY source
seems much more powerful and flexible than just supporting COPY FROM
VIEW.

Not to be a sour apple or anything but I don't see how any of this is
needed in the backend since we can easily use Psql to do it, or pretty
much any other tool.

There is an important difference between a capability in the backend vs
one synthesized in the frontend.

Not to be a sour apple or anything but I don't see how any of this is
needed in the backend since we can easily use Psql to do it, or pretty
much any other tool.

There is an important difference between a capability in the backend vs
one synthesized in the frontend.

And that would be? The suspense is killing me :)

Joshua D. Drake

--

=== The PostgreSQL Company: Command Prompt, Inc. ===
Sales/Support: +1.503.667.4564 || 24x7/Emergency: +1.800.492.2240
Providing the most comprehensive PostgreSQL solutions since 1997
http://www.commandprompt.com/

Mark Woodward wrote:

Allow COPY to output from views
Another idea would be to allow actual SELECT statements in a COPY.

Personally I strongly favor the second option as being more flexible
than the first.

I second that - allowing arbitrary SELECT statements as a COPY source
seems much more powerful and flexible than just supporting COPY FROM
VIEW.

Not to be a sour apple or anything but I don't see how any of this is
needed in the backend since we can easily use Psql to do it, or pretty
much any other tool.

There is an important difference between a capability in the backend vs
one synthesized in the frontend.

To be clear, I am not suggesting that we implement COPY TO *in* the
client. I am suggesting I don't see the purpose of COPY TO when we can
just use the psql query capability to achieve the same thing.

In talking with Andrew he is saying there is overhead in pulling these
types of result sets through the client but people that are using this
feature... won't they be connecting remotely anyway ?

I mean... why are we using COPY TO on the same server? what purpose does
it serve? What I see COPY TO for is an fast way to pull out bulk reports
or something like that... all of which I am NOT going to do on the server.

Joshua D. Drake

--

=== The PostgreSQL Company: Command Prompt, Inc. ===
Sales/Support: +1.503.667.4564 || 24x7/Emergency: +1.800.492.2240
Providing the most comprehensive PostgreSQL solutions since 1997
http://www.commandprompt.com/

VIEW.

Not to be a sour apple or anything but I don't see how any of this is
needed in the backend since we can easily use Psql to do it, or pretty
much any other tool.

There is an important difference between a capability in the backend vs
one synthesized in the frontend.

After much patience from Andrew, I withdrawn my objection :) please
continue as if I never popped my head out of my hole.

Sincerely,

Joshua D. Drake

--

=== The PostgreSQL Company: Command Prompt, Inc. ===
Sales/Support: +1.503.667.4564 || 24x7/Emergency: +1.800.492.2240
Providing the most comprehensive PostgreSQL solutions since 1997
http://www.commandprompt.com/

Tino Wildenhain <tino@wildenhain.de> writes:

Tom Lane wrote:

You're not seriously suggesting we reimplement evaluation of WHERE clauses
on the client side, are you?

no, did I? But what is wrong with something like:

\COPY 'SELECT foo,bar,baz FROM footable WHERE baz=5 ORDER BY foo' TO
file|stdout

which would just run the query (in the backend of course) and
format the output just like copy would...

Oh, I see what you have in mind. But that's hardly free either. It
requires psql to be able to translate between the frontend SELECT data
format and COPY format; which is far from a trivial thing, especially
when you consider all those CSV options ;-). And then we get to
maintain that code in parallel with the backend code anytime someone
wants another COPY feature. And then if you want the feature in a
different client, you get to do it all over again.

On balance I think doing it in the backend is more flexible and requires
less duplication of code.

regards, tom lane

I was also vaguely pondering whether all the DDL commands could be
generalized
to receive or send COPY formatted data for repeated execution. It would
be
neat to be able to prepare an UPDATE with placeholders and stream data
in COPY
format as parameters to the UPDATE to execute it thousands or millions of
times without any protocol overhead or network pipeline stalls.

MySQL already does this for INSERT :
INSERT INTO x (a,b) VALUES (1,2), (3,4), (5,6)...;

allowing arbitrary SELECT statements as a COPY source
seems much more powerful and flexible than just supporting COPY FROM
VIEW.

MySQL already does this :
SELECT INTO OUTFILE blah FROM table...

Now in both cases the MySQL syntax sucks but it's still quite practical,
and the INSERT saves some overhead (parsing, acquiring locks...) and is
quite a bit faster than regular INSERT.

PFC <lists@peufeu.com> writes:

I was also vaguely pondering whether all the DDL commands could be
generalized to receive or send COPY formatted data for repeated execution.
It would be neat to be able to prepare an UPDATE with placeholders and
stream data in COPY format as parameters to the UPDATE to execute it
thousands or millions of times without any protocol overhead or network
pipeline stalls.

MySQL already does this for INSERT :
INSERT INTO x (a,b) VALUES (1,2), (3,4), (5,6)...;

Does MySQL really let you stream that? Trying to do syntax like that in
Postgres wouldn't work because the parser would try to build up a parse tree
for the whole statement before running the command.

--
greg

Greg Stark <gsstark@mit.edu> writes:

PFC <lists@peufeu.com> writes:

MySQL already does this for INSERT :
INSERT INTO x (a,b) VALUES (1,2), (3,4), (5,6)...;

Does MySQL really let you stream that? Trying to do syntax like that in
Postgres wouldn't work because the parser would try to build up a parse tree
for the whole statement before running the command.

A quick look at MySQL's grammar doesn't show any indication that they
don't build a full parse tree too.

The above syntax is SQL-standard, so we ought to support it sometime,
performance benefits or no. I agree it might be tricky to avoid eating
an unreasonable amount of memory for a very long list in Postgres :-(

Supporting VALUES only in INSERT would be relatively trivial BTW,
but the spec actually thinks it should be allowed as a <table spec>
in FROM ...

regards, tom lane

MySQL already does this for INSERT :
INSERT INTO x (a,b) VALUES (1,2), (3,4), (5,6)...;

Does MySQL really let you stream that? Trying to do syntax like that in
Postgres wouldn't work because the parser would try to build up a parse
tree
for the whole statement before running the command.

Hehe, I don't know, but I suppose it's parsed in one-shot then executed,
and not streamed, because :
- you can append modifiers at the end of the statement (IGNORE...),
- mysql barfs if the complete SQL including data is larger than the query
buffer specified in the config file.

The second point leads to an interesting fact, ie. dumps generated by
phpmyadmin and mysqldump need a parameter specifying how long, in bytes,
the insert commands can be ; so that hopefully they can be reloaded later.
If one of the inserted values violates a "constraint", it is substituted
by "some other default value".

Still, it's useful ; and one interesting part is that everything happens
in the same SQL command (wrt concurrency).

Mark Woodward wrote:

Mark Woodward wrote:

...

This runs completely in the background and can serve as a running
backup.

And you are sure it would be much faster then a server local running
psql just dumping the result of a query?

No I can't be sure of that at all, but .... The COPY command has a
specific use that is understood and an operation that is separate from the
normal query mechanism.

Unless you change it to actually execute a query ;)

(And you could more easy avoid raceconditions in contrast to several
remote clients trying to trigger your above backup )

Again, the examples may not have been precise in presenting "why," the
focus was mostly "what" so it could be discussed. As a generic feature it
has many potential uses. Trying to debate and defend a specific use limits
the potential scope of the feature.

Thats why I'm asking. I'm still wondering which use-case actually
defends the integration of the resultset-formatter into the backend
vs. just doing it in the frontend (in both places there are already
some routines which could be used to implent).

Why have COPY anyway? Why not just use "SELECT * FROM TABLE?"

Because the special SELECT * FROM TABLE can be optimized aparently.

Ah yes, and if usual result fetch requires storing result set
in server ram, it would be nicer to change that if possible.
I think we run SELECT ... much more often then COPY ;-)
(And I hope nobody comes up with the idea if copy would be implemented
to execute queries, to generally use COPY instead of select for
large result sets in applications. Goodbye portability...)

Tom Lane <tgl@sss.pgh.pa.us> writes:

Supporting VALUES only in INSERT would be relatively trivial BTW,
but the spec actually thinks it should be allowed as a <table spec>
in FROM ...

How does that syntax work?

INSERT INTO x (a,b) from select x,y,z from t from select x2,y2,z2 from t

? doesn't seem to be very sensible?

--
greg

Greg Stark <gsstark@mit.edu> writes:

Tom Lane <tgl@sss.pgh.pa.us> writes:

Supporting VALUES only in INSERT would be relatively trivial BTW,
but the spec actually thinks it should be allowed as a <table spec>
in FROM ...

How does that syntax work?

If you look at SQL92, INSERT ... VALUES is actually not a direct
production in their BNF. They define <insert statement> as

<insert statement> ::=
INSERT INTO <table name>
<insert columns and source>

<insert columns and source> ::=
[ <left paren> <insert column list> <right paren> ]
<query expression>
| DEFAULT VALUES

<insert column list> ::= <column name list>

and then one of the alternatives for <query expression> is
<table value constructor>, which is

<table value constructor> ::=
VALUES <table value constructor list>

<table value constructor list> ::=
<row value constructor> [ { <comma> <row value constructor> }... ]

(Another alternative for <query expression> is <query specification>,
which accounts for the INSERT ... SELECT syntax.) The interesting
point here is that a <subquery> is defined as a parenthesized
<query expression>, which means that you ought to be able to use a
parenthesized VALUES list anyplace you could use a parenthesized SELECT.
So FROM lists, IN clauses, = ANY and friends, etc all really ought to be
able to support this. (A quick look at mysql's grammar suggests that
they don't handle those cases.)

The trouble with supporting it for any case other than INSERT is that
you have to work out what the column datatypes of the construct ought
to be. This is the same as the equivalent problem for UNION constructs,
but the UNION type resolution algorithm looks kinda ugly for thousands
of inputs :-(

regards, tom lane

Tom Lane wrote:

Greg Stark <gsstark@mit.edu> writes:

PFC <lists@peufeu.com> writes:

MySQL already does this for INSERT :
INSERT INTO x (a,b) VALUES (1,2), (3,4), (5,6)...;

The above syntax is SQL-standard, so we ought to support it sometime,
performance benefits or no. I agree it might be tricky to avoid eating
an unreasonable amount of memory for a very long list in Postgres :-(

Supporting VALUES only in INSERT would be relatively trivial BTW,
but the spec actually thinks it should be allowed as a <table spec>
in FROM ...

Can we just start with the simple case?

cheers

andrew

Tom Lane <tgl@sss.pgh.pa.us> writes:

The interesting point here is that a <subquery> is defined as a
parenthesized <query expression>, which means that you ought to be able to
use a parenthesized VALUES list anyplace you could use a parenthesized
SELECT. So FROM lists, IN clauses, = ANY and friends, etc all really ought
to be able to support this.

That's actually pretty neat. I've occasionally had to write queries with the
idiom

SELECT ...
FROM (SELECT a,b,c UNION ALL
SELECT d,e,f UNION ALL
SELECT g,h,i
)
WHERE ...

That's pretty awful. It would have been awfully nice to do be able to do

SELECT ... FROM (VALUES (a,b,c),(d,e,f),(g,h,i))

The trouble with supporting it for any case other than INSERT is that
you have to work out what the column datatypes of the construct ought
to be. This is the same as the equivalent problem for UNION constructs,
but the UNION type resolution algorithm looks kinda ugly for thousands
of inputs :-(

I always thought UNION just decided on the type based on the first branch and
then coerced all the others to that type. I always cast all the columns on the
first union branch just in case.

--
greg

Greg Stark wrote:

It would have been awfully nice to do be able to do

SELECT ... FROM (VALUES (a,b,c),(d,e,f),(g,h,i))

The trouble with supporting it for any case other than INSERT is that
you have to work out what the column datatypes of the construct ought
to be. This is the same as the equivalent problem for UNION constructs,
but the UNION type resolution algorithm looks kinda ugly for thousands
of inputs :-(

I always thought UNION just decided on the type based on the first branch and
then coerced all the others to that type. I always cast all the columns on the
first union branch just in case.

Could we get away with requiring an explicit type expression where
there's some abiguity or uncertainty, like this

SELECT ... FROM (VALUES (a,b,c),(d,e,f),(g,h,i)) as (a int, b text, c float)

That's what you have to do with an SRF that returns a SETOF RECORD in the same situation, after all.

cheers

andrew

In article <44804377.8070706@dunslane.net>,
Andrew Dunstan <andrew@dunslane.net> writes:

Mark Woodward wrote:

Tom had posted a question about file compression with copy. I thought
about it, and I want to through this out and see if anyone things it is a
good idea.

Currently, the COPY command only copies a table, what if it could operate
with a query, as:

COPY (select * from mytable where foo='bar') as BAR TO stdout

Isn't this already being worked on? The TODO list says:

Allow COPY to output from views

IIRC Karel Zak posted a patch for that.

Another idea would be to allow actual SELECT statements in a COPY.

Personally I strongly favor the second option as being more flexible
than the first.

How so? I see that

psql -h somehost somedb -c "copy 'somequery' to stdout" >localfile

would be more terse than

psql -h somehost somedb -c "create temp view tmp as somequery; copy tmp to stdout" >localfile

but what's more flexible there?

On Tue, Jun 06, 2006 at 04:47:40PM +0200, Harald Fuchs wrote:

In article <44804377.8070706@dunslane.net>,
Andrew Dunstan <andrew@dunslane.net> writes:

Mark Woodward wrote:

Tom had posted a question about file compression with copy. I thought
about it, and I want to through this out and see if anyone things it is a
good idea.

Currently, the COPY command only copies a table, what if it could operate
with a query, as:

COPY (select * from mytable where foo='bar') as BAR TO stdout

Isn't this already being worked on? The TODO list says:

Allow COPY to output from views

IIRC Karel Zak posted a patch for that.

Another idea would be to allow actual SELECT statements in a COPY.

Personally I strongly favor the second option as being more flexible
than the first.

How so? I see that

psql -h somehost somedb -c "copy 'somequery' to stdout" >localfile

would be more terse than

psql -h somehost somedb -c "create temp view tmp as somequery; copy tmp to stdout" >localfile

but what's more flexible there?

Flexibility aside, doing this via a temporary view is a very
non-intuitive way to go about it. AFAIK CREATE TEMP VIEW is also regular
DDL, which means more overhead in the system catalogs, along with more
need to vacuum.

I really fail to see why we shouldn't support copying from a query
unless there's some serious technical challenge. If there was a serious
technical challange that using a temporary view solved, we should do the
work of creating the temporary view for the user.
--
Jim C. Nasby, Sr. Engineering Consultant jnasby@pervasive.com
Pervasive Software http://pervasive.com work: 512-231-6117
vcard: http://jim.nasby.net/pervasive.vcf cell: 512-569-9461