Обсуждение: json indexing and data types

Hi

As json essentially only has three basic data types, string, int, and
boolean, I wonder how much of this - to index, search, and sort on
unstructured data -  is possible. I guess part of the answer would be
'jsquery and vodka', but let me describe the problem first.

The basics is, that I have a column with what is essentially json data;
a number of data structures of different depths. Perhaps 10 - 30 top
levels, and probably no more than 3, max 4 levels deep. In total there
are some hundred thousands of rows in each table. It would probably be
best stored as jsonb. Right now it's text, because it's only used by the
application itself.

It would be incredibly useful to add an index to this column, and to be
able to search, using the index, on arbitrary elements. This part seems
already there, with jsquery.

The hard part is that some of the data items really have another type.
There are dates and floating points, as the most important ones. And the
really hard part is that sorting and range searches are important,
especially for these two types. Having dates is iso-format, and
left-padding floats with zeros is a low tech solution, and especially
the latter is not very efficient.

The solution might be to add functional indexes for these data items,
but it's cumbersome and not easily maintainable. If a one-stop solution
is in the works, or already there, it could save a lot of time.

/kaare

On 12/2/15 12:03 AM, Kaare Rasmussen wrote:
>
> The hard part is that some of the data items really have another type.
> There are dates and floating points, as the most important ones. And the
> really hard part is that sorting and range searches are important,
> especially for these two types. Having dates is iso-format, and
> left-padding floats with zeros is a low tech solution, and especially
> the latter is not very efficient.
>
> The solution might be to add functional indexes for these data items,
> but it's cumbersome and not easily maintainable. If a one-stop solution
> is in the works, or already there, it could save a lot of time.

We have a client that has a similar (though also a bit different) need.
Specifically, they get an XML document that has element attributes that
tell you what data type the element should contain. We convert the XML
to JSON (easy thanks to plpython), which produces a bunch of nested JSON
objects (typed as specifically as possible in JSON). The XML attributes
get turned into items in an object. So

<some-element type="integer">42</some-element>

becomes something like

"some-element": { "@type": "integer", "#text": 42 }

Some transforms are applied to that (like replacing - with _), and the
resulting JSON is used to create a set of tables, where each table
contains one level of nesting (triggers to handle inserts are also
created). Finally, views that break out each element value are created
on top of these tables. If specific type info is available that's used
to determine the type of the column, otherwise an appropriate type is
chosen based on json_typeof(). This results in a view that looks
something like

SELECT ((json_data-> 'some_element')->>'#text')::integer AS some_element

The reason we went with one table per level was to allow full indexing
(without needing Vodka), because it made the code easier to develop (at
least during initial conception), and because it supports joining
between the views nicely (something we needed). You could probably do
this without splitting into multiple tables, but I suspect it would add
significant complexity to the view creation.

I'd like to eventually open source the guts of this, but unfortunately
there's a lot of work required to get it to the point where that would
be possible. There's also some choices that were made that in retrospect
should probably be done differently. Of course if someone wanted to pay
us to do that then we'll find the time ;). Short of that if someone is
really serious about helping with that effort I can start untangling
parts of this from the proprietary codebase that it's currently buried
in, but even that would be a pretty significant effort.
--
Jim Nasby, Data Architect, Blue Treble Consulting, Austin TX
Experts in Analytics, Data Architecture and PostgreSQL
Data in Trouble? Get it in Treble! http://BlueTreble.com

On Wed, Dec 2, 2015 at 12:03 AM, Kaare Rasmussen <kaare@jasonic.dk> wrote:
> Hi
>
> As json essentially only has three basic data types, string, int, and
> boolean, I wonder how much of this - to index, search, and sort on
> unstructured data -  is possible. I guess part of the answer would be
> 'jsquery and vodka', but let me describe the problem first.
>
> The basics is, that I have a column with what is essentially json data; a
> number of data structures of different depths. Perhaps 10 - 30 top levels,
> and probably no more than 3, max 4 levels deep. In total there are some
> hundred thousands of rows in each table. It would probably be best stored as
> jsonb. Right now it's text, because it's only used by the application
> itself.
>
> It would be incredibly useful to add an index to this column, and to be able
> to search, using the index, on arbitrary elements. This part seems already
> there, with jsquery.
>
> The hard part is that some of the data items really have another type. There
> are dates and floating points, as the most important ones. And the really
> hard part is that sorting and range searches are important, especially for
> these two types. Having dates is iso-format, and left-padding floats with
> zeros is a low tech solution, and especially the latter is not very
> efficient.
>
> The solution might be to add functional indexes for these data items, but
> it's cumbersome and not easily maintainable. If a one-stop solution is in
> the works, or already there, it could save a lot of time.

I feel your pain.  jsquery is superb for subdocument searching on *specific* subdocuments but range searching is really limited.  Value searching is there for numerics but dates and text range searching are not present.  We also have to understand that you are asking the index to make assumptions about the json that are not clear from the structure itself (such as subfield 'x' is a date).

The only workaround I've been able to come up with is to migrate the json to a specially encoded text field, stored side by side with the source json, that is more amenable to pg_trgm based searching (to give you a taste of that complexity, keys are stored upper case and values are stored lower case).

Some might say that you're better off using a dedicated json searching server like solr but these systems aren't magic; they will quickly boil down to a brute force search in the face of complex queries, and they have lots of other problems in my experience (starting with, lack of proper transactions and painfully slow insertion of large documents).  Other people recommend them; I don't.

One way of looking at this problem is that the "schemaless" check is getting cashed. If you need detailed data driven queries (as opposed to more 'test searchy' type searches) perhaps it's time to start running your data through a normalized structure.

merlin

On 12/2/15 7:06 PM, Merlin Moncure wrote:
>  > The basics is, that I have a column with what is essentially json data; a
>  > number of data structures of different depths. Perhaps 10 - 30 top
> levels,
>  > and probably no more than 3, max 4 levels deep. In total there are some
>  > hundred thousands of rows in each table. It would probably be best
> stored as
>  > jsonb. Right now it's text, because it's only used by the application
>  > itself.

After re-reading this part...

Are you in control of the JSON itself, and are the number of
permutations known in advance? It might be that something like table
inheritance is a better solution...
--
Jim Nasby, Data Architect, Blue Treble Consulting, Austin TX
Experts in Analytics, Data Architecture and PostgreSQL
Data in Trouble? Get it in Treble! http://BlueTreble.com

Merlin Moncure <mmoncure@gmail.com> writes:
> On Wed, Dec 2, 2015 at 12:03 AM, Kaare Rasmussen <kaare@jasonic.dk> wrote:
>> As json essentially only has three basic data types, string, int, and
>> boolean, I wonder how much of this - to index, search, and sort on
>> unstructured data -  is possible.

> I feel your pain.  jsquery is superb for subdocument searching on
> *specific* subdocuments but range searching is really limited.

Yeah.  The problem here is that a significant part of the argument for
the JSON/JSONB datatypes was that they adhere to standards (RFC 7159 in
particular).  I can't see us accepting a patch that changes them into
JSON-plus-some-PG-enhancements.

For cases where you know that specific sub-fields can be expected to be
of particular datatypes, I think you could get a lot of mileage out of
functional indexes ... but you'd have to write your queries to match the
indexes, which could be painful.

(Having said that, it sure looks to me like JSON's idea of a number is
float/numeric, not merely int.  Are you sure you need more capability
in that department, and if so what exactly?)

            regards, tom lane

On 2015-12-03 01:04, Jim Nasby wrote:
> We have a client that has a similar (though also a bit different)
> need. Specifically, they get an XML document that has element
> attributes that tell you what data type the element should contain. We
> convert the XML to JSON (easy thanks to plpython), which produces a
> bunch of nested JSON objects (typed as specifically as possible in
> JSON). The XML attributes get turned into items in an object. So

OK, AFAIUI, you added the schema to each row. I think that I have fewer
variations, so perhaps the information would live better outside, but
that's a detail. Turning them into tables and views is a good way to
represent the indexable data. Functionally, it seems to me to be almost
the same as functional indexing, but much more transparent, and easier
to write a query for,

> Are you in control of the JSON itself, and are the number of
> permutations known in advance? It might be that something like table
> inheritance is a better solution...

Yes, I can alter the db specification. Not sure how table inheritance
would help, though?

/kaare

On 2015-12-03 02:06, Merlin Moncure wrote:
> I feel your pain.  jsquery is superb for subdocument searching on
> *specific* subdocuments but range searching is really limited.  Value
> searching is there for numerics but dates and text range searching are
> not present.  We also have to understand that you are asking the index
> to make assumptions about the json that are not clear from the
> structure itself (such

I thought that text range searching (using indexes) was available in pg,
at least with vodka ?
> Some might say that you're better off using a dedicated json searching
> server like solr but these systems aren't magic; they will quickly
> boil down to a brute force search in the face of complex queries, and
> they have lots of other problems in my experience (starting with, lack
> of proper transactions and painfully slow insertion of large
> documents).  Other people recommend them; I don't.

They come with their own set of problems. Including not being able to be
part of a where clause. The json data may not be the only thing you want
to limit your selection with.

/kaare

On Wed, Dec 2, 2015 at 10:04 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> (Having said that, it sure looks to me like JSON's idea of a number is
> float/numeric, not merely int.  Are you sure you need more capability
> in that department, and if so what exactly?)

Numeric range searching is good, but the numeric case isn't compelling
for me.  Text string searching *is* compelling but is not as of yet
supported by jsquery.  Arbitrary token searching would be ideal, but
I'd settle for left to to right matching.

Historically, GIN index searching with over jsquery with ranges or
pg_trgm has also given me heartburn with its "worst case" performance
behavior in that it could in depressingly common cases underperform
(sometimes grossly) brute force.   This doesn't mesh well with the
'user supplied list of search terms' type of searching that we do a
lot of.

The situation of late for pg_trgm has gotten drastically better with
the triconsistent API optimizations.  I haven't gotten around yet to
seeing if any of that magic has been sprinkled on jsquery.

merlin

On Wed, Dec 2, 2015 at 7:11 PM, Jim Nasby <Jim.Nasby@bluetreble.com> wrote:
> On 12/2/15 7:06 PM, Merlin Moncure wrote:
>>
>>  > The basics is, that I have a column with what is essentially json data;
>> a
>>  > number of data structures of different depths. Perhaps 10 - 30 top
>> levels,
>>  > and probably no more than 3, max 4 levels deep. In total there are some
>>  > hundred thousands of rows in each table. It would probably be best
>> stored as
>>  > jsonb. Right now it's text, because it's only used by the application
>>  > itself.
>
>
> After re-reading this part...
>
> Are you in control of the JSON itself, and are the number of permutations
> known in advance? It might be that something like table inheritance is a
> better solution...

Yeah, or other data storage strategies that are btree friendly but not
'schema rigid', like EAV (perhaps a modified variant to support
storing the document structure).  There are definitely tradeoffs
involved but you have to consider all the options.

merlin

On 2015-12-03 05:04, Tom Lane wrote:
> Yeah. The problem here is that a significant part of the argument for
> the JSON/JSONB datatypes was that they adhere to standards (RFC 7159
> in particular). I can't see us accepting a patch that changes them
> into JSON-plus-some-PG-enhancements.

Would be nice for my specific need, but probably wouldn't do a lot of
good in the long run.

> (Having said that, it sure looks to me like JSON's idea of a number is
> float/numeric, not merely int. Are you sure you need more capability
> in that department, and if so what exactly?)

Hmm, I think you're right, having just tried some conversions. But
still, I would have to rely on vodka (or similar) knowing this, I guess?

/kaare

On 12/2/15 10:38 PM, Kaare Rasmussen wrote:
> On 2015-12-03 01:04, Jim Nasby wrote:
>> We have a client that has a similar (though also a bit different)
>> need. Specifically, they get an XML document that has element
>> attributes that tell you what data type the element should contain. We
>> convert the XML to JSON (easy thanks to plpython), which produces a
>> bunch of nested JSON objects (typed as specifically as possible in
>> JSON). The XML attributes get turned into items in an object. So
>
> OK, AFAIUI, you added the schema to each row. I think that I have fewer
> variations, so perhaps the information would live better outside, but
> that's a detail. Turning them into tables and views is a good way to
> represent the indexable data. Functionally, it seems to me to be almost
> the same as functional indexing, but much more transparent, and easier
> to write a query for,

I didn't add the schema; in this case the schema was always the same. If
you had a limited number of schemas you could indicate which one was in
a particular document and use the appropriate decoding.

>> Are you in control of the JSON itself, and are the number of
>> permutations known in advance? It might be that something like table
>> inheritance is a better solution...
>
> Yes, I can alter the db specification. Not sure how table inheritance
> would help, though?

They provide a means where you can refer to the common parts of
disparate schemas in one place, while being able to deal with the inner
details on each child table.

It might not be useful depending on what your goals are. I mentioned it
because I think most people only think of inheritance as "That weird
thing that partitioning uses."
--
Jim Nasby, Data Architect, Blue Treble Consulting, Austin TX
Experts in Analytics, Data Architecture and PostgreSQL
Data in Trouble? Get it in Treble! http://BlueTreble.com

On Wed, Dec 2, 2015 at 8:04 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Merlin Moncure <mmoncure@gmail.com> writes:
>> On Wed, Dec 2, 2015 at 12:03 AM, Kaare Rasmussen <kaare@jasonic.dk> wrote:
>>> As json essentially only has three basic data types, string, int, and
>>> boolean, I wonder how much of this - to index, search, and sort on
>>> unstructured data -  is possible.
>
>> I feel your pain.  jsquery is superb for subdocument searching on
>> *specific* subdocuments but range searching is really limited.
>
> Yeah.  The problem here is that a significant part of the argument for
> the JSON/JSONB datatypes was that they adhere to standards (RFC 7159 in
> particular).  I can't see us accepting a patch that changes them into
> JSON-plus-some-PG-enhancements.
>
> For cases where you know that specific sub-fields can be expected to be
> of particular datatypes, I think you could get a lot of mileage out of
> functional indexes ... but you'd have to write your queries to match the
> indexes, which could be painful.

If you create a view which has columns defined according to the index
expression, it does remove a lot of the pain of making queries that
use those expressions.  It looks just like using a real column, as
long as you don't update it.

Cheers,

Jeff

This is known problem, that's why we stop developing jsquery and are working on sql-level query language for jsonb, then you'll use all power and extendability of SQL.  The idea is to use power of subselects and unnest to unroll jsonb to sql level.

There is presentation at pgconf.eu on this
https://wiki.postgresql.org/images/4/4e/Createam.pdf, see slide #27

This is very interesting. Thanks for the update. And to all who answered this topic, sorry for awoling. I just got busy, but thanks for all the replies, I got something to think about.

But I'm afraid it'll come to 9.6.

I'll hope it comes in 9.6. I'll definitely look forward to that.

/kaare