Hi all,
As both developper and user of Orekit, I would prefer option 2: it has several advantages (easiest implementation, backwards-compatible), and I can't find any major drawback.
Best regards,
Pascal
Le 06/07/2018 à 16:34, MAISONOBE Luc a écrit :
Hi all,
There is currently a discussion ongoing on the Orekit developers mailing lists
about a proposed change. The decision on what to do exactly depends a lot on
how users that develop applications on top of Orekit use it, so the discussion
should really happen on this list rather than remain on the developers list.
The change is related to error handling. Currently, Orekit has one main
exception, OrekitException. This is a checked exception, so callers (i.e.
you when you develop an application on top of the library) have to either
let it bubble upward up to the main or have to catch it and act accordingly.
As most of Orekit methods declare to throw this exception, most of calling
code as to do this choice: declare or catch.
User code can also sometimes wrap up some Orekit code in other frameworks
and have to respect the API of these frameworks. In this case, the
OrekitException must always be caught and probably wrapped within a
RuntimeException so it is not "seen" by the framework. One typical example
is the use of Java 8 streams and lambda functions, and these do not allow
exception, so some cumbersome wrapping code has to be set up. Here is an
example Yannick Jeandroz sent to start the discussion on the developers
list. As the addMeasurement in the orbit determination estimator can
throw an OrekitEception, he had to write this:
iMeasurements.stream().forEach(measurement -> {
try {
estimator.addMeasurement(measurement);
}
catch (OrekitException e) {
throw new RuntimeException(e);
}
});
If addMeasurement did not throw a checked exception, this would have
been sufficient, and much more easier to read and maintain:
Measurements.stream().forEach(measurement -> estimator.addMeasurement( measurement));
So the point was that OrekitException, as a checked exception, is cumbersome.
Then several options arose during the discussion.
1) The initial proposal was that many exception arise from data loading (the dreaded
exception about TAI-UTC history, but also Sun ephemerides, Earth Orientation Parameters
and so on). When this happen, it means the caller did not set up the data configuration
correctly and nothing can be done. This should rather throw an unchecked exception and
the application should be stopped as soon as possible. This implies some work for
the Orekit team to sort out the changes to do.
2) Another proposal was to consider that all Orekit exceptions could be unchecked and
the basic contract would be: if you use Orekit, you should expect almost all code
to throw OrekitException, so at the topmost level, you should be ready to catch it
but we will not enforce it. This is the simplest solution, it almost only implies
changing
public class OrekitException extends Exception
into
public class OrekitException extends RuntimeException
3) A third proposal was to replace some OrekitException with standard Java exception
(typically IOException or ParseException for data loading).
4) In parallel to the above, another idea was that as we would change exceptions,
we should create a few different exceptions for different errors. This implies
a lot of work and would imply not going to the extrem we knew some years ago
with Apache Commons Math that was removed when we switched to Hipparchus
5) a counter-proposal to 3) was to use only a small Orekit hierarchy, thus allowing
people to catch the top-level Orekit exception and be sure everything is covered
as everything would extend this exception, this implies some work for the Orekit team
6) a slightly different version of idea 4) would be to have only 2 or 3 exceptions,
maybe one checked, one runtime, one error, this seems to be relatively easy to
do
So the point now is as top level users will see these exceptions are the one that will
need to deal with them, users should have the last word on what should be done. So
I would like to start a poll on the list to know what users prefer. It is possible
to mix several ideas. I will start by putting here the current status from the
developers mailing list. Please add your name and preferences. If you want more
insights, you can read the thread on the developers list, which is public.
Yannick : 1 + 6, not opposed to 2
Luc : 2 + 5 + 6
Evan : leaning towards just 2, though not opposed 1 + 6
Guilhem : ? + 3
Anne-Laure : 1 + 4 + 5 (3 is fine but in a "nice to have" way)
best regards,
Luc