Setting up and mastering radius search in PostgreSQL (9.1)

Radius search in PostgreSQL may come in employing a light and/or a much more sophisticated version. This article discusses the light one, namely the cube and the earth distance extensions, most probably sufficient for the web user’s getting here and there requirements. earth distance, depending on cube, assumes the earth to be perfectly spherical, anyone demanding a higher accuracy level, especially for the mountainous parts, may take a look at the PostGIS project.

Although radius search, the light variety, will be up fast and performing well, there may be some mantrap around, for the ones who prefer to read documentation the easy way too. First of all, PostgreSQL: Documentation: 9.1: earthdistance indicates that the point-based earth distance calculation is hard-wired to statute miles in units. You may use this circumstance to your advantage, like datachomp did in Radius Queries in Postgres, as long as you know what you’re doing. Second to that, taking on the alternate cube-based earth distance calculation, the earth_box function, accepting a lat/long and a radius on input, may return locations farther than the actual radius given (documented alike). This is because earth_box, as the name implies, still handles a box geometry on the idealized sphere (and not some higher order circle surface). But more on that below.



Flashback version query and the proper use of timestamp and scn clauses

Flashback version query essentially enables you to lookup the incarnations of a row (defined by primary key) in the past, in a consecutive manner. Version information is depicted by a couple of pseudo-columns, namely versions_xid, versions_startscn, versions_endscn, versions_starttime, versions_endtime and versions_operation. See Using Oracle Flashback Version Query in the docs for explanations.

In combination with flashback query or flashback transaction query, one may restore a row incarnation from the past into a new table or even rollback to a past row incarnation within the same table.

This article will discuss flashback version query together with flashback query to restore one to many rows, just shown for a row of a unique key here for brevity, detailing when and when not to use timestamp and scn select where clauses to prevent pitfalls. An example table / dataset will be given, representing a real world scenario where some past data needs to be identified first and is then to be made available again.

Flashback version query uses the following pattern, including the pseudo-columns introduced above, on an actual application-, but not a system-table (alike flashback transaction query). A timestamp– or scn-range must be supplied to define the lookup window (defined by the stock of the available undo-data, remember) and to actually populate the pseudo-columns, respectively:

SELECT pseudo_col1, ..., app_col1, ... FROM app_tab
  { SCN | TIMESTAMP } { expr | MINVALUE }
  AND { expr | MAXVALUE } ]

Flashback query then provides for the historic data selection in terms of application-table columns, typically used to run a create table as select..., if you like, supplying an as of timestamp– or scn selector:

-- CTAS, maybe
SELECT app_col1, ... FROM app_tab
  AS OF { SCN | TIMESTAMP } expr

All theory so far, yet the real example upfront, make shure have grasped what this trinity in fact means:

  • Flashback version query
  • Flashback query
  • Flashback transaction query

Also be prepared to unhesitantly get down to pen and paper when it comes to history lookup time windows and row level incarnations of data. It’s quite tricky and bewildering sometimes, be concentrated, have a sketch of timelines at hand (you know for https://en.wikipedia.org/wiki/Primer_(film), for example, some visualization like https://en.wikipedia.org/wiki/File:Time_Travel_Method-2.svg may come in very handy).


Grepping the actual bind values of sql through v$sql_bind_capture

Programmatically or intercepted by forced cursor sharing, sql where clause actual values shall always be bound to variable placeholders and never / no longer be literally coded. You know Tom Kyte talking that story on each and every occasion, pointing out that so many sql areas got stuck in size and performance because of literal values and the resulting multitude of variants of sql statements. On the other hand, people complain about the sporadic issue that the optimizer takes wrong decisions due to one-time probing of actual values in a given session, a nightmare in 10g, much improved by adaptive cursor sharing throughout 12c meanwhile. Well, I do not want to dive into this again, what follows is just a commented recipe sql statement to inspect what actual bind values have had been in action for a (potentially) sub-optimal sql execution. Please note, that the STATISTICS_LEVEL initialization parameter takes to be greater than BASIC to have the statement deliver any data.

The statement is neither complex nor long-running, just using two perfomance views v$sql_bind_capture and v$sqlarea, resp. While v$sql_bind_capture provides for the bind names, values, capture state and time, v$sqlarea offers different ways to approach the subject, by schema or sql-id or module etc, and not at least, the raw text of the statement in question.


Getting a raw constant number of rows from oracle’s table sample function

You may of course know these two famous posts called To sample or not to sample… (part-2) about data sampling by Mark Hornick. Although very limited in scope, the two posts (imho) very well sketch why we may employ data sampling and how we may lift off table sampling in oracle.
In general, sampling is used to make a representative statement about a collection of data while only regarding a limited random selection, the sample. As long as you are ok to analyze just a sufficient subset of your 1o million rows table for an analysis, you will save your environment a lot of resources and time. On some other scenario, a limited random data selection may also serve verification or testing purposes where, however, not the representativeness but the randomness at a more or less constant sample size, determines the quality of the sample output. Again, as long as you are ok to not exceed this 15 minutes time window overnight, you will be allowed to run that live unit test on any table in question, on 1, 10 or 100 million rows.
In sql, selecting in regard to gain a representative statement will feed the sample function with a requested percentage of rows to sample from. This is what the oracle sample function already offers. Yet another sql to accept a requested actual number of rows to return, independent of the table size, is not available so far (although most people do expect exactly this behaviour when they spot the sql sample function for the first time, weird). The following text will outline a snippet of pl/sql to provide for a sample function to accept the expected number of rows as a parameter.


XMLType.schemaValidate is your only true friend for xmlschema validation

When it comes to introducing xml-data into your database storage, you might, as all sincere developers do, at first attempt to take care of proper data integrity checking on import. Since xml is a really powerful but a somewhat complex document-alike data representation, such integrity checking must incorporate proving the document strucuture, as sort of integrated data types, iff you like, as well as the document data, in terms of facets of given actual values. According to the xml standard, xmlschema is the means of choice here, offering another xml-spelled specification layer to achive the two beforementioned goals. Eventually, a given xml-instance will have to be thrown against the xmlschema provided, to assure its integrity (and being well-formed too, btw), the earlier the better, at best on data import already.

The paragraph above is actually not very oracle specific. Any implementation of the xml standard, as oracle’s xml db does, proposes this course of action. However, since some xml operations may become quite costly when the xml-instances get large and the whole dom tree has to be set up in memory, oracle, as others, dabbles at dodging and shifting pricey work to the most reasonable extend possible. You may load xml into the database, claiming any xml-instances are just fine, validate xml-instances only to be well-formed, manage the validation status of xml-instances at your own responsibility and so forth. There is nothing wrong about that whatsoever. The point is, though, it might be a stony walkway to learn to distinguish the maybe from the certain, to establish a reliable check to safeguard your xml data integrity, just as simple as ALTER TABLE mytab ADD(CONSTRAINT mytab_CHK CHECK(INSERT_TS is not null) NOT DEFERRABLE ENABLE);.

Jep, what next? I’m going to briefly discuss the xmlschema / xmlinstance used below and then show what I experienced will happen using the various ways of “validating” the instance against the schema. The post is somewhat lengthy, but do not get frightened, this is for the c&p examples all along the way.