By understanding how real users have employed reliable multicast in real distributed systems, we can develop insight concerning the degree to which this technology has matched expectations. This paper reviews a number of applications with that goal in mind. Our findings point to trade-offs between the form of reliability used by a system and its scalability and performance. We also find that to reach a broad user community (and a commercially interesting market) the technology must be better integrated with component and object-oriented systems architectures. Looking closely at these architectures, however, we identify some assumptions about failure handling which make reliable multicast difficult to exploit. Indeed, the major failures of reliable multicast are associated with attempts to position it within objectoriented systems in ways that focus on transparent recovery from server failures. The broader opportunity appears to involve relatively visible embeddings of these tools into object-oriented architectures enabling knowledgeable users to make trade-offs. Fault-tolerance through transparent server replication may be better viewed as an unachievable holy grail. stock markets into sophisticated networks. The maturing of reliable multicast technology is a critical gating factor in enabling such a path. The Swiss ExchangeThe Swiss Exchange [5] is an ambitious all-electronic stock exchange: the physical trading floor has been supplanted by a collection of workstations running exchange-developed software situated in member bank and brokerage trading rooms. The old physical trading floor now houses the development and management teams which run the electronic system. The benefits to exchange members are several: rather than calling in orders to traders on the floor, order processing occurs in the same room where institutional trading strategy is developed. Trading support tools can thus be exploited without revealing proprietary methods, and the floor representative of the bank can participate directly in strategy discussions without being observed by competitors.The task of the software implementing the Swiss Exchange is to distribute to each participating trader every trade and every quoted price in the identical order, fault-tolerantly, throughout the full set of perhaps 1000 active traders at 60 member banks and institutions. The latency of the architecture must be low (at worst, a few seconds), and data rates peak at approximately 45 events per second, each represented by a message in the 512-1024 byte range. The exchange guarantees fairness to its customers by giving each exactly the same information, in exactly the same order, and at nearly the same time. The basic architecture of the exchange system is hierarchical: a server at the core sees all trades and all quotes; the latter dominate by about 20:1. It broadcasts these to hubs located at each of the participating institutions; they, in turn, relay the information to the workstations of the traders. Reliable multicast enters this picture in two ways. It...