It looks like you haven’t weighted your analysis; the increase you are seeing is the result of an increase in ACS sample size between 2004 and 2005. You need to use weights to get correct point estimates for sample data; this blog post offers a quick overview of sampling weights. There is not a change to the underlying occupation codes between these two years. I see an increase in the unweighted counts of about 5,000 people with an OCC2010 value of 4610 and 13,000 with an OCC2010 value of 3600; this seems like it matches the figure you shared. However, when I apply weight the cross-tab using PERWT, the abrupt shift disappears: