InvenSense gyros tend to drift pretty badly/fast, and setting a low-pass down-sampling filter only gets you so far.
A slower magnetometer is important as it usually offers the only true consistent directional reference using magnetic declination (once calibrated). Unfortunately, these "9-axis" units are not immune to problems like large magnets (speakers can pin a sensor into error) or large metal objects.
Generally, people will combine the sensor-fusion steps with a multi-axis Kalman filter, and just use Quaternion math to avoid singularities. Thus, you end up weighting the certainty of each 3-axis sensor, suppressing drift/noise, and handle the ambiguities caused in a "6-axis" sensor.
It really depends on what you are trying to build, and who you have around to help you....
The chip maker may be more useful in this regard.... I am still traumatized by a terrible job I had getting these chips on spec... hint: the error bound increases non-linearly with the angle offset, and can be a pain to tune for a prior-state dependent filter...
Best of luck, =)