Quick SELECT sometimes time out - sql

Question

I have stored procedure which execute simple select. Any time I run it manually, it runs under the second. But in production (SQL Azure S2 database) it runs inside scheduled task every 12 ours - so I think it is reasonable to expect it to run every time with "cold" - with no cached data. And the performance is very unpredictable - sometimes it takes 5 second, sometimes 30 and sometimes even 100.
The select is optimized to the maximum (of my knowledge, anyway) - I created filtered index including all the columns returned from SELECT, so the only operation in execution plan is Index scan. There is huge difference between estimated and actual rows:
But overall the query seems pretty lightweight. I do not blame environment (SQL Azure) because there is A LOT of queries executing all the time, and this one is the only one with this performance problem.
Here is XML execution plan for SQL ninjas willing to help : http://pastebin.com/u5GCz0vW
EDIT:
Table structure:
CREATE TABLE [myproject].[Purchase](
[Id] [int] IDENTITY(1,1) NOT NULL,
[ProductId] [nvarchar](50) NOT NULL,
[DeviceId] [nvarchar](255) NOT NULL,
[UserId] [nvarchar](255) NOT NULL,
[Receipt] [nvarchar](max) NULL,
[AppVersion] [nvarchar](50) NOT NULL,
[OSType] [tinyint] NOT NULL,
[IP] [nchar](15) NOT NULL,
[CreatedOn] [datetime] NOT NULL,
[ValidationState] [smallint] NOT NULL,
[ValidationInfo] [nvarchar](max) NULL,
[ValidationError] [nvarchar](max) NULL,
[ValidatedOn] [datetime] NULL,
[PurchaseId] [nvarchar](255) NULL,
[PurchaseDate] [datetime] NULL,
[ExpirationDate] [datetime] NULL,
CONSTRAINT [PK_Purchase] PRIMARY KEY CLUSTERED
(
[Id] ASC
)
Index definition:
CREATE NONCLUSTERED INDEX [IX_AndroidRevalidationTargets3] ON [myproject].[Purchase]
(
[ExpirationDate] ASC,
[ValidatedOn] ASC
)
INCLUDE ( [ProductId],
[DeviceId],
[UserId],
[Receipt],
[AppVersion],
[OSType],
[IP],
[CreatedOn],
[ValidationState],
[ValidationInfo],
[ValidationError],
[PurchaseId],
[PurchaseDate])
WHERE ([OSType]=(1) AND [ProductId] IS NOT NULL AND [ProductId]<>'trial' AND ([ValidationState] IN ((1), (0), (-2))))
Data can be considered sensitive, so I cant provide sample.

Answer 1

Since your query returns only 1 match, I think you should trim down your index to a bare minimum. You can get the remaining columns via a Key Lookup from the clustered index:
CREATE NONCLUSTERED INDEX [IX_AndroidRevalidationTargets3] ON [myproject].[Purchase]
(
[ExpirationDate] ASC,
[ValidatedOn] ASC
)
WHERE ([OSType]=(1) AND [ProductId] IS NOT NULL AND [ProductId]<>'trial' AND ([ValidationState] IN ((1), (0), (-2))))
This doesn't eliminate the scan, but it makes the index much leaner for a fast read.
Edit: OP stated that the slimmed-down index was ignored by SQL Server. You can force SQL Server to use the filter index:
SELECT *
FROM [myproject].[Purchase] WITH (INDEX(IX_AndroidRevalidationTargets3))