It seems that the problem comes from the fact that all the available connexions are busy, and especially from the fact that several customer who are connected almost permanently.
Therefore, we are going to increase the number of available connexions per server and increase also the number of servers in the cluster to avoid the congestion risks.