Diff for "Database/Performance"

Not logged in - Log In / Register

Differences between revisions 8 and 9
Revision 8 as of 2011-09-27 12:23:54
Size: 8271
Editor: adeuring
Comment:
Revision 9 as of 2011-09-27 12:27:54
Size: 8272
Editor: adeuring
Comment:
Deletions are marked like this. Additions are marked like this.
Line 111: Line 111:
    SELECT * FROM Table WHERE column1 > value1 AND column2 > value2     SELECT * FROM Table WHERE (column1, column2) > (value1, value2)

Poor query times - looks right, takes ages

Normally, the simplest form of the query is the fastest.

Checklist of known problems

  1. subselects, EXISTS, IS IN - these can be useful for optimizing particular edge cases, but can degrade performance for the normal cases.
  2. database views. These used to be useful to keep our code sane, but it is now clearer to use Storm to express the view logic from the client side. Database views now serve little purpose for us except to hide unnecessary joins that will degrade performance.
  3. bad query plans generated on the DB server - talk to a team lead to get an explain analyze done on staging, or to Stuart or a LOSA to get the same done on production (if the staging one looks ok its important to check production too).
    • Bad plans can happen due to out of date statistics or corner cases, sometimes rewriting the query to be simpler/slightly different can help. Specific things to watch out for are nested joins and unneeded tables (which can dramatically change the lookup).
  4. fat indices - if the explain looks sensible, talk to Stuart about this.
  5. missing indices - check that the query should be indexable (and if in doubt chat to Stuart).
  6. using functions in ORDER BY: calling functions on every row of an intermediary table - if a sort cannot be answered by iterating an index then postgresql will generate an intermediate table containing all the rows that match the constraints, and sort that in-memory; functions that affect the sort order have to be evaluated before doing the sort - and thus before any LIMIT occurs.
  7. Querying for unrelated tables. Quite possibly either prejoins, or prepopulation of derived attributes. Look for a code path that is narrower, or pass down a hint of some sort about the data you need so the actual query can be more appropriate. Sometimes more data is genuinely needed but still messes up the query: consider using a later query rather than prejoining. E.g. using the pre_iter_hook of DecoratedResultSet to populate the storm cache.

Many small queries [in a webapp context] -- Late Evaluation

Databases work best when used to work with sets of data, not objects - but we write in python, which is procedural and we define per-object code paths.

One particular trip up that can occur is with related and derived data.

Consider:

def get_finished_foos(self):
    return Store.of(self).find(Foo, And(Foo.me == self.id, Foo.finished == True))

This will perform great for one object, but if you use it in a loop going over even as few as 30 or 40 objects you will cause a large amount of work - 30 to 40 separate round trips to the database.

Its much better to prepopulate a cache of these finished_foos when you request the owning object in the first place, when you know that you will need them.

To do this, use a Tuple Query with Storm, and assign the related objects to a cached attribute which your method can return. For attributes the @cachedproperty('_foo_cached') can be used to do this in combination with a DecoratedResultSet

Be sure to clear these caches with a Storm invalidation hook, to avoid test suite fallout. Objects are not reused between requests on the appservers, so we're generally safe there. (Our storm and sqlbase classes within the Launchpad tree have these hooks, so you only need to manually invalidate if you are using storm directly).

A word of warning too - Utilities will often get in the way of optimising this :)

Diagnosis Tools and Approaches

EXPLAIN ANALYZE on staging and qastaging can be used by LOSAs, the TAs, and squad leads.

If you want to see how a query is working on a GET page locally, try the ++oops++ and ++profile++ tools. ++profile++ reportedly works on staging and qastaging now too.

Unfortunately, they sometimes do not work properly for POSTs, and can't be used in other scenarios. See 641969, for instance.

If you are working on a test and want to see how a query is working, try one of these tools.

  • Use the built-in Storm debug tracer. If you start with this...
       def test_some_storm_code(self):
           < some setup logic >
           < the Storm-using code I'm curious about >
           < more stuff >
  • ..then you can use the debug tracer to see what's going on. When you run your tests after changing the code to look like the below, stdout will include the queries run, and timestamps for start and finish.
       def test_some_storm_code(self):
           < some setup logic >
           from storm.tracer import debug; debug(True)
           try:
               < the Storm-using code I'm curious about >
           finally:
               debug(False)
           < more stuff >
  • StormStatementRecorder, LP_DEBUG_SQL=1, LP_DEBUG_SQL_EXTRA=1, QueryCollector. In extremis you can also turn on statement logging in postgresql. [Note: please add more detail if you are reading this and have the time and knowledge.]

  • Raise an exception at a convenient point, to cause a real OOPS.

Efficient batching of SQL result sets: StormRangeFactory

Batched result sets are rendered via the class canonical.launchpad.webapp.bachting.BatchNavigator. (This class is a thin wrapper around lazr.batchnavigator.BatchNavigator.)

StormRangeFactory delegates the retrieval of batches from a result set to an IRangeFactory (defined in lazr.batchnavigator.interfaces); The default range factory is lazr.batchnavigator.ListRangeFactory.

This factory uses regular Python slicing to access a batch, which is mapped by the ORM to a query like

    SELECT ... FROM ... OFFSET o LIMIT l;

for a slice operation result_set[o:o + l].

Calculating offsets can be very expensive for result sets with a large number of rows. StormRangeFactory uses a different approach: Given a query

    SELECT * FROM Table ORDER BY Table.column1, Table.column2;

and given a batch where the values of column1, column2 in last row of the batch are value1, value2, it generates a query for the next batch by adding a WHERE clause:

    SELECT * FROM Table WHERE (column1, column2) > (value1,  value2)
        ORDER BY Table.column1, Table.column2 LIMIT batchsize+1;

Usage

The main change to use StormRangeFactory is simple: Just replace

    batchnav = BatchNavigator(resultset, request)

with

    from canonical.launchpad.webapp.batching import StormRangeFactory
    batchnav = BatchNavigator(
        resultset, request, range_factory=StormRangeFactory(resultset))

Limitations

StormRangeFactory needs access to the columns used for sorting; it retrieves the values of the sort columns automatically, using the resultset.order_by() parameters. This has several consequences:

1. Result sets must be entire model objects:

    store.find(Person)
    store.order_by(Person.id)

can be used with StormRangeFactory, but

    store.find(Person.id)
    store.order_by(Person.id)

can not be used.

2. The order_by parameters must be Storm Column instances, optionally wrapped into Desc().

    resultset.order_by(Person.id)
    resultset.order_by(Desc(Person.id))

works, but

    resultset.order_by('Person.id')

does not work.

3. Obviously, all sort column must appear in the result set. This means that

    resultset = store.find(
        BugTask, BugTask.product == Product, Product.project == context)
    resultset.order_by(Product.name, BugTask.id)

does not work. Use

    resultset = store.find(
        (BugTask, Product), BugTask.product == Product,
        Product.project == context)
    resultset.order_by(Product.name, BugTask.id)

instead and wrap this result set into DecoratedResultSet.

4. StormRangeFactory works only with regular Storm ResultSets and with DecoratedResultSets, but not with legacy SQLObjectResultSets.

    resultset = store.find(Person)
    resultset.order_by(Person.id)

works, but

       resultset = Person.select(...)

does not work.

Database/Performance (last edited 2011-09-28 09:06:12 by adeuring)