An acceptable (at least it was where I worked) method of testing capacitors is to use a "C" meter, or any of the newer digital multimeters having a capacitance test function and compare readings with the rating and tolerance of the cans, rejecting those out of tolerance. Of course you have to convert the kvar ratings into microfarads. You can also use a constant voltage source such as 120 volts AC and read current through the capacitor, again converting kvar into current at the test voltage.
The capacitors with higher voltage ratings are constructed of several parallel capacitor packs arranged in series in the can. Different manufacturers of the same kvar and voltage units quite often use totally different arrangements. Failure of a can seems to be preceded by an unnoticed failure of at least one pack, usually a short. When a pack becomes shorted the capacitance of the can increases. If you know what the arrangement is you can actually calculate how much the readings will increase when a pack fails.
I discovered several years ago that the actual tolerance of a large number of units from the same manufacturer was actually a LOT smaller than the published tolerance. I believe the packs, in the higher voltage units anyway, are wound by machine and are extremely consistent. So what I always liked to do instead of looking at ratings and tolerances is calculate the mean and standard deviation of the cans in a capacitor bank being tested. Consider good all units reading the mean plus or minus 3 standard deviations. I picked 3 because, if I remember correctly, for a Normal distribution only about 1 reading in 10,000 will fall outside plus or minus 3 standard deviations due to manufacturing variations. You will find that higher voltage units will mostly read mean plus or minus 1 standard deviation. The only cans to be concerned with are those reading extremely low or a little high. The published ratings and internal construction are such that in many cases cans with at least one shorted pack can have capacitance readings that fall within tolerance. While my method is certainly not very scientific and my old statistics prof would probably scream, it works. To see this method in action just take a large number of capacitor unit readings and put them in a bar chart. All the readings will fall below about mean plus about 2+ standard deviations. Then there will be some readings above mean plus 3 standard deviations. Those are the bad ones, having at least one shorted pack.
After finding the mean and std. dev. of a particular manufacturers capacitor unit I would use the same readings from then on. The readings never change until a pack fails. Then the unit's readings will get it rejected.
Now for the real world part. I have found that my simple little system is good for finding units with shorted packs and are doomed to fail, but are still within the manufacturers tolerance. But many times I have seen capacitor banks containing hundreds of cans to be tested one day and have failures the next day. No other method I ever saw was any better. Some manufacturers seem to have much worse failure rates than others. The only thing I ever saw that improved the operating record is to use units with a much higher voltage rating than applied voltage, but then you pay the price in loss of kvars. Oh well.
