We study the population properties of merging binary black holes in the second LIGO--Virgo Gravitational-Wave Transient Catalog assuming they were all formed dynamically in gravitationally bound clusters. Using a phenomenological population model, we infer the mass and spin distribution of first-generation black holes, while self-consistently accounting for hierarchical mergers. Considering a range of cluster masses, we see compelling evidence for hierarchical mergers in clusters with escape velocities $gtrsim 100~mathrm{km,s^{-1}}$. For our most probable cluster mass, we find that the catalog contains at least one second-generation merger with $99%$ credibility. We find that the hierarchical model is preferred over an alternative model with no hierarchical mergers (Bayes factor $mathcal{B} > 1400$) and that GW190521 is favored to contain two second-generation black holes with odds $mathcal{O}>700$, and GW190519, GW190602, GW190620, and GW190706 are mixed-generation binaries with $mathcal{O} > 10$. However, our results depend strongly on the cluster escape velocity, with more modest evidence for hierarchical mergers when the escape velocity is $lesssim 100~mathrm{km,s^{-1}}$. Assuming that all binary black holes are formed dynamically in globular clusters with escape velocities on the order of tens of $mathrm{km,s^{-1}}$, GW190519 and GW190521 are favored to include a second-generation black hole with odds $mathcal{O}>1$. In this case, we find that $99%$ of black holes from the inferred total population have masses that are less than $49,M_{odot}$, and that this constraint is robust to our choice of prior on the maximum black hole mass.