4.2Availability management (availability rate, redundancy, single point of failure)

When a service manager wants to raise availability, blindly lining up expensive equipment yields nothing. Availability is determined by two factors—"resistance to breaking (reliability)" and "ease of being fixed (maintainability)"—and the availability of the whole system varies greatly with how the components are connected (series versus parallel redundancy). This section quantifies the availability rate as MTBF / (MTBF + MTTR) and, using that value as a clue, builds judgment to diagnose and design where the weak point (single point of failure) is and whether redundancy or shorter repair time is the more effective lever. The calculations center not on "the value itself" but on "what the value tells you to improve."

4.2.1Availability rate, MTBF (reliability), and MTTR (maintainability)

• The MTBF (mean time between failures) is the average uptime from one failure to the next and represents reliability (resistance to breaking). The MTTR (mean time to repair) is the average time from failure to completion of repair and represents maintainability (ease of being fixed). The availability rate = MTBF / (MTBF + MTTR)—the fraction of total time during which the service was usable normally.
• There are two directions for raising availability: lengthen the MTBF (adopt high-reliability parts, do preventive maintenance, reduce load to cut failures themselves) or shorten the MTTR (monitoring, automatic recovery, spare parts, and rapid response by maintenance staff to fix faster). For example, with MTBF = 990 hours and MTTR = 10 hours, availability is 990 / (990 + 10) = 990 / 1000 = 0.99 (99%). If the MTTR is halved from 10 to 5 hours, it becomes 990 / 995 ~ 0.995—availability improves through repair speed alone without changing failure frequency—so judge by the situation which is more cost-effective to improve.

4.2.2Overall availability of series and parallel-redundant configurations

• A series configuration is one where the whole stops if any single element fails (the service does not hold unless all components are normal). Overall availability is the product of each element's availability: a1 x a2 x .... For example, connecting two elements of availability 0.99 in series gives 0.99 x 0.99 = 0.9801 (about 98.01%)—the more elements added in series, the lower the overall availability. Here the lowest-availability element tends to become the single point of failure that drags down the whole.
• In parallel redundancy, several elements with the same function are provided, and the service continues as long as at least one is alive. Overall availability is 1 - (1 - a)^n (a = each element's availability, n = the number in parallel), subtracting from 1 the probability that all elements fail simultaneously. For example, two elements of availability 0.9 in parallel give 1 - (1 - 0.9)^2 = 1 - 0.1^2 = 1 - 0.01 = 0.99 (99%). For elements of availability 0.99, two in parallel give 1 - 0.01^2 = 1 - 0.0001 = 0.9999 (99.99%)—redundancy raises overall availability substantially, but at the cost of duplicated equipment and maintenance.

4.2.3Diagnosing and eliminating a single point of failure (SPOF)

• A single point of failure (SPOF) is a spot whose failure stops the entire system. Even if servers are made redundant with two units, if both share a single power supply, a single network switch, or a single load balancer, that shared part remains as a SPOF. Redundancy is meaningful not as "apparent duplication" but only when every path by which a failure propagates to a total outage is duplicated, without exception.
• Eliminating a SPOF is best done not by "making everything redundant" but by prioritizing the paths with the largest business impact, which matters from a cost-effectiveness standpoint. Using the availability calculation (estimating the whole as a product of series terms) to identify which element drags overall availability down the most and adding parallel redundancy there yields the highest return on investment. Conversely, over-duplicating an element that contributes little to overall availability only piles up cost.

Suppose the service manager of an e-commerce site is diagnosing a situation in which the site cannot reliably meet the SLA-defined availability of 99.9% (about 8.76 hours of annual downtime), with brief outages recurring a few times a month. Decomposing the configuration as a series reveals a web server (availability 0.999), an application server (0.999), a database (0.995), and shared storage (0.99) connected in series, so overall availability is estimated as the product 0.999 x 0.999 x 0.995 x 0.99. Approximating this product gives about 0.983, and the lowest terms—the shared storage (0.99) and the database (0.995)—are dragging down the whole, meaning these are the single points of failure. Deciding to "uniformly double every server to two units" here is an inefficient error: making the web or application servers (already 0.999) redundant improves overall availability only slightly and wastes the investment. The correct judgment is to add parallel redundancy to the lowest-availability shared storage. Making the 0.99 storage a two-unit parallel redundancy raises that part's availability to 1 - (1 - 0.99)^2 = 1 - 0.0001 = 0.9999, removing the overall-availability bottleneck. In addition, if failure frequency (MTBF) cannot be lowered, one can also shorten the MTTR—strengthening monitoring and using automatic failover to cut the time from detection to switchover. The trap here is a situation of "we made the servers redundant with two units, yet availability did not rise," where in fact the two units share a single power supply or single storage, so the shared part remained as a SPOF. A service manager uses availability figures as a tool for identifying bottlenecks and judges to add redundancy first to the weakest path with the largest business impact.

4.2.4Section summary

• Availability = MTBF / (MTBF + MTTR); MTBF is reliability (resistance to breaking), MTTR is maintainability (ease of being fixed)
• Series overall availability is the product of availabilities (lower as you add more); parallel redundancy is 1 - (1 - a)^n (higher as you add more)
• A single point of failure remains at the weakest series element or a shared part—add redundancy first to the path with the largest business impact