Why is it a problem that your response variable is skewed in regression? Is taking logarithms the only way to solve it?
Taking logarithms is not a one-size-fits-all approach - what if your response is negatively skewed?
"Simple" linear regression assumes that the response variable is normally distributed. If you'd like to perform this kind of regression, then you should probably transform your response variable to be as normal as possible.
There are other approaches / model types to deal with skewed responses - check out the Tweedie family.