Hi Michael,
I'm sorry that my initial response to your question about JDO-50 was
delayed in arriving on jdo-dev. I will try to remember to include your
name on the two list in the future to make sure that you receive a
response asap.
Anyway, as I said, this fix is probably unnecessary because existing
values in the double_values array cover the same functional case as is
proposed by this issue. Also, at this point it is more worthwhile to
put our time into getting existing tests to run and writing tests for
JDO 2 functionality rather than improving the tck11 tests. So I think
you should add your ideas as a comment to the issue in case we decide to
pursue it in the future and close it as "won't fix". It's interesting
to know that this was not as easy as it looked.
Thanks,
Michelle
Michael Watzek wrote:
Hi,
I like to propose a fix for JIRA JDO-50.
This issue suggests to add two floating point numbers to the TCK query
operator tests: 0.1 and -0.1. On the first glance, this change seems
to be trivial. However, it is not a trivial change. For this reason, I
like to discuss the fix beginning with a brief summery of the general
concept, and the implementation strategy of the query operator tests:
The general concept:
Class AllTypes maintains arrays for all supported types. The values in
these arrays represent an initial state of the database. There are
query tests for all JDOQL query operators. Each of these tests rely on
the initial database state. They perform queries on all supported
types checking the result against expected values.
The current implementation:
- The arrays of all supported types have a certain length (10).
- 10 objects of class AllTypes are made persistent. This is the
initial database state.
- Each quey test performs more or less 30 separate queries on each
supported type, and additionally the same amount of queries for
wrapper types. As a very rough estimation, there may be 2000-3000
separate queries in the operators package.
- Each of these queries is passed a constant values as parameter and
the expected query result which depends on the certain array length
(10). The parameters must match values in the initial arrays of all
supported types.
Adding floating point values 0.1 and -0.1 to the implementation means:
- Increasing each of the initial arrays by two elements. Regardless,
if this makes sense or not (e.g. boolean arrays).
- Adapting the expected query result of all separate queries (about
2000 or 3000). This action has to be taken because the expected query
result depends on the certain array length (10).
- Adding new float and double queries for parameter values 0.1 and -0.1.
Obviously, this is a very big change for a very little functionality
enhancement. For this reason, I propose to change the implementation
strategy of the concept with the objective to make the implementation
independent from the initial database state. If we gained this
independency, we could extend the initial arrays without changing the
query operator tests.
The proposal:
In a first step, we try to reach independeny of the fixed array
lengths. Furthermore, we try to be independent of the constant values
for query parameters and the expected query results. After this step
has been taken, all arrays still have the same length (as now), but we
do not have 2000-3000 separate query calls in the code. Instead, we
have loops for each supported type in each query test. Inside each
loop, we have exactly one query call. The loops iterate over the
elements of the initial arrays. The parameter values are given by the
current array element. The expected query result can be derived by the
current array index.
The amount of query calls in the code will dramatically decrease. On
the other hand, we can be sure that all values of the initial arrays
are queried which is not the case today. This means that we gain
completeness for the query operator tests. Another advantage is the
extensibility of the implementation wrt. the initial arrays.
In a second step (which we can postpone to the future), we allow the
initial arrays to have different sizes. Thus, a boolean array would
have two elements (or three if we would have NULL tests).
I believe that the effort to implement the proposal does not exceed
the effort to add the floating points the the current implementation.
It may take about 4-5 days without step two.
Please let me know what you think.
Thanks,
Michael
