Обсуждение: BLOB

> Hello
>
> I would like to know
> - how to create under psql a BLOB field in sql language
> - how store a value (and not a file) in sql language
> - the maximum size of the BLOB field
   For  now  (and  7.0),  there  is no way to have rows >8K in a   regular table.
   But I'm actually making alot  of  progress  in  breaking  it.   Seems we'll get rid of this restriction in 7.1.
   To share some info with the other developers:
   I  just  successfully  stored the entire PG sources (~10M) in   one table.  Stored are all 936  .[chyl]  files.
The tables   schema is:
 
       sources (         s_path    text,         s_data    clob       )
   clob  is  a  user  defined,  very  simple varsize datatype, I   created for testing. The entire source  tree
inserts in  23   seconds.  If  I  cut  off all sources at 7K and insert into a   text field it needs 8 seconds.  Not
bad,because  the  amount   of data shrinks <4M due to the cutoff.
 
   Reading the data back shows similar results, 0.8 secs vs. 0.3   secs.
   And the toaster already uses a  heap+index  to  store  values   external.
   This  all done with full toasting - i.e. try to compress, and   only  if  it  still  doesn't  fit  move  already
compressed  attribute  into secondary relation. As a side effect, the 10M   of sources only need 3.4M main and 2.6M
secondaryheap.
 
   I  experienced  some  trouble  with   occational   "Deadlock"   detections,  when  using  multiple processes
hammeringon the   toaster. But I was able to reproduce the same  error  with  a   plain  7.0  and  "text"  attributes
(onlyfar less frequent).   Seems there's  something  wrong  in  our  deadlock  detection   algorithm.
 
   Another  thing  is,  that  ISTM  that index's never shrink on   vacuum. One requirement of TOAST  is,  that  the
secondaries  tables  index-OID is remembered in the external reference, so   the ref can be  passed  around  in  the
entire backend  and   whenever needed, it's real value can be fetched quickly by an   index scan. Thus, it's not
possibleto drop/recreate an index   on that right now. Seems we need a "REBUILD INDEX" utility or   the like.
 
   Anyway, TOAST is on it's way. And as soon as we go  for  7.1,   I'll  need  alot  of  help  to  make  all  our
existingtypes   toastable, add  administrative  utility  commands  and  teach   pg_dump to deal with all that.
 


Jan

--

#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me.                                  #
#========================================= wieck@debis.com (Jan Wieck) #

wieck@debis.com (Jan Wieck) writes:
>     To share some info with the other developers:
>     I  just  successfully  stored the entire PG sources (~10M) in
>     one table.  Stored are all 936  .[chyl]  files.

Cool!  Sounds like you are making great progress.

>     Anyway, TOAST is on it's way. And as soon as we go  for  7.1,
>     I'll  need  alot  of  help  to  make  all  our existing types
>     toastable, add  administrative  utility  commands  and  teach
>     pg_dump to deal with all that.

Question: pg_dump shouldn't need to know about toasted tuples as
such, should it?  I'm guessing that what you are thinking of is
adding administration capabilities to control whether particular
columns are toast-able, and then pg_dump would need to know about
saving and loading those control fields.  But toasting a particular
field value ought to be invisible to applications...
        regards, tom lane

Tom Lane wrote:

> wieck@debis.com (Jan Wieck) writes:
> >     To share some info with the other developers:
> >     I  just  successfully  stored the entire PG sources (~10M) in
> >     one table.  Stored are all 936  .[chyl]  files.
>
> Cool!  Sounds like you are making great progress.
   At least I think so :-)


> >     Anyway, TOAST is on it's way. And as soon as we go  for  7.1,
> >     I'll  need  alot  of  help  to  make  all  our existing types
> >     toastable, add  administrative  utility  commands  and  teach
> >     pg_dump to deal with all that.
>
> Question: pg_dump shouldn't need to know about toasted tuples as
> such, should it?  I'm guessing that what you are thinking of is
> adding administration capabilities to control whether particular
> columns are toast-able, and then pg_dump would need to know about
> saving and loading those control fields.  But toasting a particular
> field value ought to be invisible to applications...
   Answer:  Right!  The  only  change from a clients PoV is that   attribute values can be bigger now.
   So it's only to teach pg_dump to correctly  restore  all  the   possible  tuning.   For  now  there  is  an
attstorage(char)   attribute telling in the TupleDesc, what can be done  with  a   particular value in a HeapTuple.
Themeanings are:
 
   'p'   don't  touch, either DBA doesn't want or the type isn't         capable of dealing with it.   'e'   never try
tocompress, if needed store external as  is.   'x'   try  to compress and if needed store compression result
external.  'm'   try to keep in main  tuple  as  long  as  possible,  if         needed handle like 'x'.
 
   If  a  tuple doesn't fit into a block, the toaster performs 4   loops over the attributes, each one  stops
immediately when   the tuple fits.
 
   1.  Inline compress storage 'x' attributes.
   2.  Move   storage   'x'  and  'e'  attributes  to  secondary       relation.
   3.  Inline compress storage 'm' attributes.
   4.  Move storage 'm' attributes to secondary relation.
   My plan now is to add a default storage value to pg_type.  At   CREATE  TABLE  time, a secondary relation plus it's
indexare   automatically  created  if  attributes  of  the   table   are   toastable.  This is questionable, because
it'llcost 2 inodes   and 16K per table, but  we  can  change  it  later  and  have   special  admin  commands to deal
withit. The toaster doesn't   jump in if a relation doesn't have a secondary  one.  But  it   will  expand values from
otherrelations in that case (INSERT   ... SELECT).
 
   With the pg_type entry telling what to do by default, we  can   work on all our types one by one, and enable them at
thetime   all their type specific functions can handle it.
 
   Finally, I'd like to give  the  secondary  relations  another   relkind,  so they are ignored by vacuum, and then
vacuumthem   explicitly while the LOCK on the master table is still  held.
 


Jan

--

#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me.                                  #
#========================================= wieck@debis.com (Jan Wieck) #

> -----Original Message-----
> From: pgsql-sql-owner@hub.org [mailto:pgsql-sql-owner@hub.org]On Behalf
> Of Jan Wieck
> 
> > Hello
> >
> > I would like to know
> > - how to create under psql a BLOB field in sql language
> > - how store a value (and not a file) in sql language
> > - the maximum size of the BLOB field
> 
>     For  now  (and  7.0),  there  is no way to have rows >8K in a
>     regular table.
> 
>     But I'm actually making alot  of  progress  in  breaking  it.
>     Seems we'll get rid of this restriction in 7.1.
> 
>     To share some info with the other developers:
> 
>     I  experienced  some  trouble  with   occational   "Deadlock"
>     detections,  when  using  multiple processes hammering on the
>     toaster. But I was able to reproduce the same  error  with  a
>     plain  7.0  and  "text"  attributes (only far less frequent).
>     Seems there's  something  wrong  in  our  deadlock  detection
>     algorithm.
>

Does the trouble mean that "Deadlock" message was issued for
non-deadlock cases ? 
>     Another  thing  is,  that  ISTM  that index's never shrink on
>     vacuum. One requirement of TOAST  is,  that  the  secondaries
>     tables  index-OID is remembered in the external reference, so
>     the ref can be  passed  around  in  the  entire  backend  and
>     whenever needed, it's real value can be fetched quickly by an
>     index scan. Thus, it's not possible to drop/recreate an index
>     on that right now. Seems we need a "REBUILD INDEX" utility or
>     the like.
>

REINDEX command could do it in a sense.
However current implementation couldn't guarantee the existence
of  proper index in case of abort. Proper implementation would
require the standard transactional control mechanism to handle
relation files such as new naming rule of relation files....

Regards.

Hiroshi Inoue
Inoue@tpf.co.jp 

Jan Wieck writes:

>     I  just  successfully  stored the entire PG sources (~10M) in
>     one table.  Stored are all 936  .[chyl]  files.

Cool, then we can get rid of CVS. :-)

>     clob  is  a  user  defined,  very  simple varsize datatype, I
>     created for testing.

Keep it, it's SQL3.

>     Seems there's  something  wrong  in  our  deadlock  detection
>     algorithm.

Our deadlock detection "algorithm" is that when nothing happens for 1 sec
then that's a deadlock. Increasing that number might make those messages
go away but that's still far from an algorithm, of course.

>     I'll  need  alot  of  help  to  make  all  our existing types
>     toastable,

I'm wondering how transparent all of this will be. What is involved in
making existing types toastable? How does that affect user defined
datatypes now and in the future?


-- 
Peter Eisentraut                  Sernanders väg 10:115
peter_e@gmx.net                   75262 Uppsala
http://yi.org/peter-e/            Sweden

Peter Eisentraut <peter_e@gmx.net> writes:
>> I'll  need  alot  of  help  to  make  all  our existing types
>> toastable,

> I'm wondering how transparent all of this will be. What is involved in
> making existing types toastable? How does that affect user defined
> datatypes now and in the future?

I'd like to think that we can fold the TOAST support into the
argument-access macros that I plan to be defining for the fmgr rewrite.
See previous rants on this subject, eg 14-Jun-99, 23-Oct-99 (in
pgsql-hackers archives).

Net result for user-defined-datatype authors will be "if you revise
your routines, they will be easier to read, more portable, and will
support TOASTed values.  If you don't, they'll still work about as
well (or poorly) as they did before."
        regards, tom lane

Peter Eisentraut wrote:

> >     clob  is  a  user  defined,  very  simple varsize datatype, I
> >     created for testing.
>
> Keep it, it's SQL3.
   We don't need it later. At the time text is toastable, we can   simply create an alias for it and are done. And I
think blob   is bytea then, isn't it?
 

> >     Seems there's  something  wrong  in  our  deadlock  detection
> >     algorithm.
>
> Our deadlock detection "algorithm" is that when nothing happens for 1 sec
> then that's a deadlock. Increasing that number might make those messages
> go away but that's still far from an algorithm, of course.
   Really?  I  remember  that  someday it was "if I'm waiting 60   seconds for my lock, ...". What a change.
   What's the exact location of that parameter to change?

> >     I'll  need  alot  of  help  to  make  all  our existing types
> >     toastable,
>
> I'm wondering how transparent all of this will be. What is involved in
> making existing types toastable? How does that affect user defined
> datatypes now and in the future?
   Let's look at the input/output functions of clob:
   void *   clob_in(char *s)   {       void       *c;       int32       l;
       if (s == NULL)           return NULL;
       l = strlen(s);       c = (void *)palloc(VARHDRSZ + l);
       VARATT_SIZEP(c) = l + VARHDRSZ;       memcpy(VARATT_DATA(c), s, l);
       return c;   }

   char *   clob_out(void *c)   {       void       *p;       char       *s;       int32       l;
       if (c == NULL)       {           s = palloc(2);           s[0] = '-';           s[1] = '\0';           return s;
     }
 
       VARATT_GETPLAIN(c, p);
       l = VARATT_SIZE(p) - VARHDRSZ;       s = (char *)palloc(l + 1);       memcpy(s, VARATT_DATA(p), l);       s[l] =
'\0';
       VARATT_FREE(c, p);
       return s;   }
   So the input function doesn't change at all.  Only  functions   that  have  a  toastable type as argument need to
wraparound   with   a   local   variable   and   the    VARATT_GETPLAIN(),   VARATT_FREE() macros.
 
   VARATT_GETPLAIN(arg,  ptr)  places  the plain value of arg in   ptr.  If the argument wasn't toasted, it's assigned
asis. If   compressed   or   stored  external,  the  original  value  is   reconstructed in palloc()'d memory and
assignedto ptr.
 
   VARATT_FREE(arg, ptr) free()'s ptr if it  is  different  from   arg.
   It  all ain't that complicated. Easy enough to use it in user   defined types  too  (a  must  because  today's  user
defined   functions  usually  use  our base types too). Only that there   are hundreds of functions in utils/adt and
contribthat  need   to be looked at.
 


Jan

--

#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me.                                  #
#========================================= wieck@debis.com (Jan Wieck) #

Tom Lane wrote:

> I'd like to think that we can fold the TOAST support into the
> argument-access macros that I plan to be defining for the fmgr rewrite.
> See previous rants on this subject, eg 14-Jun-99, 23-Oct-99 (in
> pgsql-hackers archives).
   So  we need to fold the VARATT_... macro's I'm using now into   the FMGR_... macros. Or define the macro's you want
now.

> Net result for user-defined-datatype authors will be "if you revise
> your routines, they will be easier to read, more portable, and will
> support TOASTed values.  If you don't, they'll still work about as
> well (or poorly) as they did before."
   Sorry, but that ain't true. Making an existing type toastable   means  that  all  functions,  receiving  that type
needto be   revised.
 


Jan

--

#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me.                                  #
#========================================= wieck@debis.com (Jan Wieck) #

wieck@debis.com (Jan Wieck) writes:
>> Net result for user-defined-datatype authors will be "if you revise
>> your routines, they will be easier to read, more portable, and will
>> support TOASTed values.  If you don't, they'll still work about as
>> well (or poorly) as they did before."

>     Sorry, but that ain't true. Making an existing type toastable
>     means  that  all  functions,  receiving  that type need to be
>     revised.

What I meant was that they'd still work, with a limit on field size,
just like before.  ie, no TOAST support.
        regards, tom lane

> wieck@debis.com (Jan Wieck) writes:
> >> Net result for user-defined-datatype authors will be "if you revise
> >> your routines, they will be easier to read, more portable, and will
> >> support TOASTed values.  If you don't, they'll still work about as
> >> well (or poorly) as they did before."
>
> >     Sorry, but that ain't true. Making an existing type toastable
> >     means  that  all  functions,  receiving  that type need to be
> >     revised.
>
> What I meant was that they'd still work, with a limit on field size,
> just like before.  ie, no TOAST support.
   Yes, but at the first time, a toasted value is handed to them   the result (up to backend crash)  is  unpredictable.
So  any   user   defined   function   taking   "text"  as  argument  is   potentially in danger!
 
   Better tell them they have to revise.


Jan

--

#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me.                                  #
#========================================= wieck@debis.com (Jan Wieck) #

wieck@debis.com (Jan Wieck) writes:
>> What I meant was that they'd still work, with a limit on field size,
>> just like before.  ie, no TOAST support.

>     Yes, but at the first time, a toasted value is handed to them
>     the result (up to backend crash)  is  unpredictable.  So  any
>     user   defined   function   taking   "text"  as  argument  is
>     potentially in danger!

Oh, right, user-defined functions on system types will have problems
if the system type has been marked toastable.  I was thinking of
user-defined datatypes, for which it'd be OK to leave the type
untoastable if you didn't want to fix the associated functions right
away.

>     Better tell them they have to revise.

Well, hmm.  It's not out of the question that we could continue to
support old user-defined functions even on toastable types.  There is
going to be a compatibility wrapper anyway around any function that's
adhering to the old fmgr interface.  So with just a little more work,
we could make that wrapper expand any toasted values that are being
presented to the function.  Only new-style functions would ever get
passed toasted values directly.

This'd also make it a lot less painful to convert the built-in
functions, of course.

The only downside I see in this is that it'd take a few extra catalog
lookups to determine which arguments are toastable types and thus
potentially in need of untoasting.  I don't think that'd be a big loss,
since we normally only do the catalog lookups for a function reference
once per query anyway.

Seem reasonable?
        regards, tom lane

Jan Wieck writes:

>     Really?  I  remember  that  someday it was "if I'm waiting 60
>     seconds for my lock, ...". What a change.
> 
>     What's the exact location of that parameter to change?

include/config.h:#define DEADLOCK_CHECK_TIMER 1

This only has second resolution. I changed it to milliseconds for my
configuration clean-up (details coming soon), but I'm not sure if that's
worth a lot.

The logic behind this is in backend/storage/lmgr/proc.c.

>     Only that there
>     are hundreds of functions in utils/adt and contrib that  need
>     to be looked at.

What datatypes besides text'ish and binary data'ish do you want to toast?
SQL(3) only defines CLOB and BLOB, so they might be thinking along my
lines. I mean surely having toastable polygons and numerics has some
theoretical value but is it worth bothering?


-- 
Peter Eisentraut                  Sernanders väg 10:115
peter_e@gmx.net                   75262 Uppsala
http://yi.org/peter-e/            Sweden

Peter Eisentraut <peter_e@gmx.net> writes:
> What datatypes besides text'ish and binary data'ish do you want to toast?
> SQL(3) only defines CLOB and BLOB, so they might be thinking along my
> lines. I mean surely having toastable polygons and numerics has some
> theoretical value but is it worth bothering?

Good point.  The only possible candidates for toasting are varlena
types, which are

select oid,typname from pg_type
where  typtype='b' and typlen < 0 and typname !~ '^_';oid  | typname
------+---------  17 | bytea  25 | text  32 | SET 602 | path 604 | polygon 705 | unknown 869 | inet 650 | cidr1042 |
bpchar1043| varchar1560 | bit1562 | varbit1625 | lztext1700 | numeric
 
(14 rows)

plus array types (of which there are lots, but only one set of access
routines needs to deal with toasting).  I find it hard to foresee any
need for toasted inet/cidr data ;-), and I'm not even sure what SET is.
And unknown is just a placeholder, and lztext is going away anyway.
So I'd say we have

MUSTs: bytea, text, bpchar, varchar, arrays

WANTs: bit, varbit, numeric

LOW PRIORITY: path, polygon

which is not so many types after all... although some of these are used
by a lot of routines:

select pg_type.oid,typname,count(*) from pg_type, pg_proc
where  typtype='b' and typlen < 0 and typname !~ '^_' and pg_type.oid
in (proargtypes[0],proargtypes[1],proargtypes[2],proargtypes[3],
proargtypes[4],proargtypes[5],proargtypes[6],proargtypes[7])
group by pg_type.oid,typname;oid  | typname | count
------+---------+-------  17 | bytea   |     6  25 | text    |    88 602 | path    |    29 604 | polygon |    24 869 |
inet   |    161042 | bpchar  |    221043 | varchar |    141560 | bit     |    151562 | varbit  |    151625 | lztext  |
 121700 | numeric |    45
 
(11 rows)

(This is an overestimate since routines with multiple arguments may be
counted in more than one category...)
        regards, tom lane