Temporal irregularity of the output of volcanic material is studied for the sequence of large (V ≥ 0.5 km3, N = 29) explosive eruptions on Kamchatka during the last 10,000 years. Informally, volcanic productivity looks episodic, and dates of eruptions cluster. To investigate the probable self-similar clustering behavior of eruption times, we determine correlation dimension Dc. For intervals between events 800 and 10,000 years, Dc ≈ 1 (no self-similar clustering). However, for shorter delays, Dc = 0.71, and the significance level for the hypothesis Dc < 1 is 2.5%. For the temporal structure of the output of volcanic products (i.e., for the sequence of variable-weight points), a self-similar “episodic” behavior holds over the entire range of delays 100–10,000 years, with Dc = 0.67 (Dc < 1 at 3.4% significance). This behavior is produced partly by the mentioned common clustering of event dates, and partly by another specific property of the event sequence, that we call “order clustering”. This kind of clustering is a property of a time-ordered list of eruptions, and is manifested as the tendency of the largest eruptions (as opposed to smaller ones) to be close neighbors in this list. Another statistical technique, of “rescaled range” (R/S), confirms these results. Similar but weaker-expressed behavior was also found for two other data sets: historical Kamchatka eruptions and acid layers in Greenland ice column. The episodic multiscaled mode of the output of volcanic material may be a characteristic property of a sequence of eruptions in an island arc, with important consequences for climate forcing by volcanic aerosol, and volcanic hazard.