Azure Front Door: Resiliency Series – Part 3: Tenant isolation
Abhishek Tiwari, Vice President of Engineering, Azure Networking Amit Srivastava, Partner Director of PM, Azure Networking Varun Chawla, Partner Director of Engineering, Azure Networking Azure Front Door serves hundreds of thousands of tenants from hundreds of edge locations, densely sharing the globally distributed edge fleet. That density is exactly what allows us to deliver global scale, performance, and cost efficiency. It also means that, without strong isolation, a single tenant’s incompatible configuration or anomalous traffic can, in the worst case, affect many other tenants. The October 2025 incidents reinforced how important it is to contain this class of risk. Our goal for this tenant isolation is simple to state and hard to achieve: no single tenant’s configuration or traffic should be able to impact any other tenant. In Part 1 of our three part mini blog series, we outlined our four‑pillar strategy for improving the resiliency of Azure Front Door: configuration resiliency, data plane resiliency, tenant isolation, and accelerated Recovery Time Objective (RTO). Part 1 detailed how we would make configuration propagation safer and how the data plane keeps serving from a ‘last‑known‑good’ (LKG) configuration, even if an incompatible configuration change is propagated to the data plane. Part 2 turned to recovery, showing how we bring the system back to full operation in an accelerated, predictable, and in a bounded timeframe. In this final part, we turn to the tenant isolation pillar that ensures that any single tenant configuration or traffic issues are limited in scope to that tenant alone and do not impact other tenants. We will also show how Azure Front Door achieves single-tenant containment through configuration isolation, lazy loading, and a micro-cellular layered ingress-sharding architecture. Repair status: all outstanding items now complete Before we dive into tenant isolation, here is a final update on the overall repair items from the two October 2025 incidents (you can review the details in our Azure Incident Retrospective sessions for the October 9th and October 29th incidents). We are pleased to report that all outstanding work across every pillar is now complete and fully deployed in production – including the tenant isolation work described in this post. With these safeguards in place, we have also returned configuration propagation latency to pre‑incident levels while keeping platform stability our top priority. In the table below, “Completed” means broadly deployed in production. Learning category Goal Repairs Status Safe customer configuration deployment Incompatible configuration never propagates beyond ‘EUAP or canary regions’ Control plane and data plane defect fixes; forced synchronous configuration processing; additional stages with extended bake time; early detection of crash state Completed Data plane resiliency Configuration processing cannot impact data plane availability Manage data‑plane lifecycle to prevent outages caused by configuration‑processing defects; isolated work‑process in every data plane server to process and load the configuration Completed 100% Azure Front Door resiliency posture for Microsoft internal services Microsoft operates an isolated, independent Active/Active fleet with automatic failover for critical Azure services Phase 1: onboarded critical services batch impacted on Oct 29th outage running on a day‑old configuration; Phase 2: automation & hardening of operations, auto‑failover and self‑management of onboarding for additional services Completed Recovery improvements Data plane crash recovery in under 10 minutes Data plane boot‑up time optimized via local cache; recovery time accelerated to under 10 minutes Completed Tenant isolation No configuration or traffic regression can impact other tenants Micro‑cellular Azure Front Door with ingress layered shards Completed Why isolation at edge scale is deceptively hard Traditional isolation techniques such as dedicating separate hardware to each tenant or running every tenant inside its own virtual machine are impractical at the edge. Edge sites are constrained on space, power, and capacity. The entire premise of a modern, multi-tenant application delivery platform is that any tenant can be served from any site closest to the user. We cannot simply partition hundreds of thousands of tenants onto dedicated machines without giving up either proximity, scale, or the efficiency that make the edge fast and cost efficient. Isolation therefore, must be achieved in software, inside a multi-tenant fleet. What we already do today Azure Front Door already includes several layers of tenant isolation and partitioning. However, the incidents in October clearly highlighted that these techniques were not enough. Prior to the incidents, our protection mechanisms included: Infrastructure partitioning. Edge sites were organized into physically isolated primary and fallback traffic rings. Noisy‑neighbor protection. Fair‑share resource allocation, rate limiting, and anomaly-based load protection kept any single tenant from monopolizing shared resources such as CPU, memory, or network bandwidth on an Azure Front Door server. Circuit breakers. Circuit breakers shed costly work first and can disable a risky per‑tenant feature before it exhausts shared resources on a server. Real‑time crash protection. A crash‑analysis system correlates crash signatures across machines and can pinpoint and block crash patterns caused by tenant IPs or traffic patterns. While these protections are valuable, many of them are reactive and proved insufficient during the October incidents. The next generation of isolation makes single‑tenant containment a fundamental part of the platform which governs how configurations are loaded, and how traffic is served. Configuration isolation: loading only what is needed Part 2 introduced ‘lazy loading’ as a recovery optimization technique. It is also an important configuration‑isolation mechanism. Historically, every worker on every edge server had to be ready to serve any tenant, which meant each worker loaded a large set of tenant configurations. A single incompatible configuration could therefore ripple across many workers. With lazy loading, a worker loads a tenant’s configuration and its TLS certificates only when it actually receives traffic for that tenant. The practical consequence for isolation is powerful: a faulty configuration can only affect the workers that have loaded that specific tenant, never the entire server or fleet. Combined with per‑tenant validation on load, and the Food Taster safeguard from Part 1 (a sacrificial process that pretests every configuration change in isolation), configuration problems are caught early and contained to the smallest possible footprint. Figure 1: With lazy loading, an incompatible configuration is contained to the workers serving that tenant, instead of poisoning the whole server. Tenant isolation: a micro-cellular, layered ingress sharding architecture Configuration isolation limits the blast radius of an incompatible configuration. Traffic isolation addresses the other half of the problem: a tenant whose anomalous traffic incident, like a sudden surge, a pathological request pattern, or malicious activity, could degrade a shared worker. Our approach is a micro‑cellular architecture that combines multiple concepts working together. Worker‑process isolation. Each edge server already runs many independent worker processes. Instead of letting every worker serve every tenant, we assign tenants to specific groups of workers. Those worker group (shards) become the unit of isolation: if a tenant destabilizes its shard, the impact is contained to that shard’s workers while the rest of the server keeps serving normally. Ingress sharding. Rather than a handful of fixed shards, we compose shards from overlapping subsets of a server’s workers. Even a modest number of workers can be combined into an enormous number of distinct, overlapping shards – giving us a very large number of fine‑grained fault domains without dedicating hardware to every tenant. Figure 2: Tenants are randomly assigned to different shards on each server (layer). Even when a good tenant shares the noisy tenant’s shard on one layer, routing steers its traffic to healthy shards on the others. Multi‑layer ingress sharding. This is where ‘layered’ comes in. Each edge server is treated as an independent layer, and each tenant is assigned to a different, randomly chosen shard on every server. Because assignments are independent from one server to the next, two tenants that happen to share a shard on one server are extremely unlikely to share a shard again on another server. The chance of any good tenant repeatedly colliding with a noisy tenant across many servers becomes vanishingly small. Intelligent ingress routing. Tying it together is a routing layer that terminates each incoming connection, identifies the tenant, and steers the request to that tenant’s assigned, healthy shard. If a shard is unhealthy or saturated, traffic is directed to the tenant’s healthy shards on other layers. Figure 3: Intelligent Ingress Routing The combined effect is that when a noisy tenant overwhelms or crashes its shard on one server, only that shard is affected. Because every other tenant is spread across a different, randomized set of shards, they continue to find healthy paths, and the routing layer moves their traffic accordingly. A worst-case availability problem is downgraded to, at most, a small and redistributable capacity problem, that the routing layer smooths over. An in-depth technical analysis of layered ingress sharding is available here for reference. Shrinking the blast radius Taken together, these mechanisms fundamentally change the shape of failure. In a uniform, fully shared fleet, an incompatible tenant can, in the worst case, affect a large share of the tenants on a machine, in an edge location, or beyond. With configuration isolation and layered ingress sharding, the same failure is confined to a subset of workers serving the offending tenant. Our target for this tenant isolation pillar is effectively single‑tenant containment: a configuration or traffic anomaly caused by one tenant should never cause issues to any other tenants. Figure 4: From a shared fleet where a single tenant can affect many, to micro‑cellular shards that confine impact to the offending tenant. Validating isolation in practice As with our recovery work, we don’t simply design these boundaries and assume they hold, we test them. Through deliberate fault‑injections, we have pushed noisy and faulty tenants into the system and confirmed that impact stayed contained to the offending shard, that healthy tenants kept serving, and that the routing layer steered around unhealthy shards as intended. This turns isolation from a design claim into a well-drilled, and repeatable outcome. Closing This post concludes our three-part mini blog series on Azure Front Door resiliency. We have shared how we are making configuration propagation safer (Part 1), recovering faster when failures do occur (Part 2), and containing the blast radius stemming from any single tenant through configuration isolation and a micro‑cellular, layered‑sharding architecture (Part 3). Resiliency, however, is not a project with an end date. It is an ongoing commitment. While this series concludes the blog series of our response to the October 2025 incidents, our investments in Azure Front Door’s resiliency, isolation, and recovery will continue. As we make further improvements, we will keep sharing them with you. We deeply value our customers’ trust in Azure Front Door. We remain committed to exceeding expectations for security, reliability, and transparency.Azure Front Door: Resiliency Series – Part 2: Faster recovery (RTO)
Abhishek Tiwari, Vice President of Engineering, Azure Networking Amit Srivastava, Partner Director of PM, Azure Networking Varun Chawla, Partner Director of Engineering, Azure Networking Karthik Uthaman, Principal Engineer, Azure Networking In Part 1 of this blog series, we outlined our four‑pillar strategy for resiliency in Azure Front Door: configuration resiliency, data plane resiliency, tenant isolation, and accelerated Recovery Time Objective (RTO). Together, these pillars help Azure Front Door remain continuously available and resilient at global scale. Part 1 focused on the first two pillars: configuration and data plane resiliency. Our goal is to make configuration propagation safer, so incompatible changes never escape pre‑production environments. We discussed how incompatible configurations are blocked early, and how data plane resiliency ensures the system continues serving traffic from a last‑known‑good (LKG) configuration even if a bad change manages to propagate. We also introduced ‘Food Taster’, a dedicated sacrificial process running in each edge server’s data plane, that pretests every configuration change in isolation, before it ever reaches the live data plane. In this post, we turn to the recovery pillar. We describe how we have made key enhancements to the Azure Front Door recovery path so the system can return to full operation in a predictable and bounded timeframe. For a global service like Azure Front Door, serving hundreds of thousands of tenants across 210+ edge sites worldwide, we set an explicit target: to be able to recover any edge site – or all edge sites – within approximately 10 minutes, even in worst‑case scenarios. In typical data plane crash scenarios, we expect recovery in under a second. Repair status The first blog post in this series mentioned the two Azure Front Door incidents from October 2025 – learn more by watching our Azure Incident Retrospective session recordings for the October 9 th incident and/or the October 29 th incident. Before diving into our platform investments for improving our Recovery Time Objectives (RTO), we wanted to provide a quick update on the overall repair items from these incidents. We are pleased to report that the work on configuration propagation and data plane resiliency is now complete and fully deployed across the platform (in the table below, “Completed” means broadly deployed in production). With this, we have reduced configuration propagation latency from ~45 minutes to ~20 minutes. We anticipate reducing this even further – to ~15 minutes by the end of April 2026, while ensuring that platform stability remains our top priority. Learning category Goal Repairs Status Safe customer configuration deployment Incompatible configuration never propagates beyond ‘EUAP or canary regions’ Control plane and data plane defect fixes Forced synchronous configuration processing Additional stages with extended bake time Early detection of crash state Completed Data plane resiliency Configuration processing cannot impact data plane availability Manage data-plane lifecycle to prevent outages caused by configuration-processing defects. Completed Isolated work-process in every data plane server to process and load the configuration. Completed 100% Azure Front Door resiliency posture for Microsoft internal services Microsoft operates an isolated, independent Active/Active fleet with automatic failover for critical Azure services Phase 1: Onboarded critical services batch impacted on Oct 29 th outage running on a day old configuration Completed Phase 2: Automation & hardening of operations, auto-failover and self-management of Azure Front Door onboarding for additional services March 2026 Recovery improvements Data plane crash recovery in under 10 minutes Data plane boot-up time optimized via local cache (~1 hour) Completed Accelerate recovery time < 10 minutes April 2026 Tenant isolation No configuration or traffic regression can impact other tenants Micro cellular Azure Front Door with ingress layered shards June 2026 Why recovery at edge scale is deceptively hard To understand why recovery took as long as it did, it helps to first understand how the Azure Front Door data plane processes configuration. Azure Front Door operates in 210+ edge sites with multiple servers per site. The data plane of each edge server hosts multiple processes. A master process orchestrates the lifecycle of multiple worker processes, that serve customer traffic. A separate configuration translator process runs alongside the data plane processes, and is responsible for converting customer configuration bundles from the control plane into optimized binary FlatBuffer files. This translation step, covering hundreds of thousands of tenants, represents hours of cumulative computation. A per edge server cache is kept locally at each server level – to enable a fast recovery of the data plane, if needed. Once the configuration translator process produces these FlatBuffer files, each worker processes them independently and memory-maps them for zero-copy access. Configuration updates flow through a two-phase commit: new FlatBuffers are first loaded into a staging area and validated, then atomically swapped into production maps. In-flight requests continue using the old configuration, until the last request referencing them completes. The data process recovery is designed to be resilient to different failure modes. A failure or crash at the worker process level has a typical recovery time of less than one second. Since each server has multiple such worker processes which serve customer traffic, this type of crash has no impact on the data plane. In the case of a master process crash, the system automatically tries to recover using the local cache. When the local cache is reused, the system is able to recover quickly – in approximately 60 minutes – since most of the configurations in the cache were already loaded into the data plane before the crash. However, in certain cases if the cache becomes unavailable or must be invalidated because of corruption, the recovery time increases significantly. During the October 29 th incident, a data plane crash triggered a complete recovery sequence that took approximately 4.5 hours. This was not because restarting a process is slow, it is because a defect in the recovery process invalidated the local cache, which meant that “restart” meant rebuilding everything from scratch. The configuration translator process then had to re-fetch and re-translate every one of the hundreds of thousands of customer configurations, before workers could memory-map them and begin serving traffic. This experience has crystallized three fundamental learnings related to our recovery path: Expensive rework: A subset of crashes discarded all previously translated FlatBuffer artifacts, forcing the configuration translator process to repeat hours of conversion work that had already been validated and stored. High restart costs: Every worker on every node had to wait for the configuration translator process to complete the full translation, before it could memory-map any configuration and begin serving requests. Unbounded recovery time: Recovery time grew linearly with total tenant footprint rather than with active traffic, creating a ‘scale penalty’ as more tenants onboarded to the system. Separately and together, the insight was clear: recovery must stop being proportional to the total configuration size. Persisting ‘validated configurations’ across restarts One of the key recovery improvements was strengthening how validated customer configurations are cached and reused across failures, rather than rebuilding configuration states from scratch during recovery. Azure Front Door already cached customer configurations on host‑mounted storage prior to the October incident. The platform enhancements post outage focused on making the local configuration cache resilient to crashes, partial failures, and bad tenant inputs. Our goal was to ensure that recovery behavior is dominated by serving traffic safely, not by reconstructing configuration state. This led us to two explicit design goals… Design goals No category of crash should invalidate the configuration cache: Configuration cache invalidation must never be the default response to failures. Whether the failure is a worker crash, master crash, data plane restart, or coordinated recovery action, previously validated customer configurations should remain usable—unless there is a proven reason to discard it. Bad tenant configuration must not poison the entire cache: A single faulty or incompatible tenant configuration should result in targeted eviction of that tenant’s configuration only—not wholesale cache invalidation across all tenants. Platform enhancements Previously, customer configurations persisted to host‑mounted storage, but certain failure paths treated the cache as unsafe and invalidated it entirely. In those cases, recovery implicitly meant reloading and reprocessing configuration for hundreds of thousands of tenants before traffic could resume, even though the vast majority of cached data was still valid. We changed the recovery model to avoid invalidating customer configurations, with strict scoping around when and how cached entries are discarded: Cached configurations are no longer invalidated based on crash type. Failures are assumed to be orthogonal to configuration correctness unless explicitly proven otherwise. Cache eviction is granular and tenant‑scoped. If a cached configuration fails validation or load checks, only that tenant’s configuration is discarded and reloaded. All other tenant configurations remain available. This ensures that recovery does not regress into a fleet‑wide rebuild due to localized or unrelated faults. Safety and correctness Durability is paired with strong correctness controls, to prevent unsafe configurations from being served: Per‑tenant validation on load: Each cached tenant configuration is validated during the ‘load and verification’ phase, before being promoted for traffic serving. Therefore, failures are contained to that tenant. Targeted re‑translation: When validation fails, only the affected tenant’s configuration is reloaded or reprocessed. Therefore, the cache for other tenants is left untouched. Operational escape hatch: Operators retain the ability to explicitly instruct a clean rebuild of the configuration cache (with proper authorization), preserving control without compromising the default fast‑recovery path. Resulting behavior With these changes, recovery behavior now aligns with real‑world traffic patterns - configuration defects impact tenants locally and predictably, rather than globally. The system now prefers isolated tenant impact, and continued service using last-known-good over aggressive invalidation, both of which are critical for predictable recovery at the scale of Azure Front Door. Making recovery scale with active traffic, not total tenants Reusing configuration cache solves the problem of rebuilding configuration in its entirety, but even with a warm cache, the original startup path had a second bottleneck: eagerly loading a large volume of tenant configurations into memory before serving any traffic. At our scale, memory-mapping, parsing hundreds of thousands of FlatBuffers, constructing internal lookup maps, adding Transport Layer Security (TLS) certificates and configuration blocks for each tenant, collectively added almost an hour to startup time. This was the case even when a majority of those tenants had no active traffic at that moment. We addressed this by fundamentally changing when configuration is loaded into workers. Rather than eagerly loading most of the tenants at startup across all edge locations, Azure Front Door now uses an Machine Learning (ML)-optimized lazy loading model. In the new architecture, instead of loading a large number of tenant configurations, we only load a small subset of tenants that are known to be historically active in a given site, we call this the “warm tenants” list. The warm tenants list per edge site is created through a sophisticated traffic analysis pipeline that leverages ML. However, loading the warm tenants is not good enough, because when a request arrives and we don’t have the configuration in memory, we need to know two things. Firstly, is this a request from a real Azure Front Door tenant – and, if it is, where can I find the configuration? To answer these questions, each worker maintains a hostmap that tracks the state of each tenant’s configuration. This hostmap is constructed during startup, as we process each tenant configuration – if the tenant is in the warm list, we will process and load their configuration fully; if not, then we will just add an entry into the hostmap where all their domain names are mapped to the configuration path location. When a request arrives for one of these tenants, the worker loads and validates that tenant’s configuration on demand, and immediately begins serving traffic. This allows a node to start serving its busiest tenants within a few minutes of startup, while additional tenants are loaded incrementally only when traffic actually arrives—allowing the system to progressively absorb cold tenants as demand increases. The effect on recovery is transformative. Instead of recovery time scaling with the total number of tenants configured on a server, it scales with the number of tenants actively receiving traffic. In practice, even at our busiest edge sites, the active tenant set is a small fraction of the total. Just as importantly, this modified form of lazy loading provides a natural failure isolation boundary. Most Edge sites won’t ever load a faulty configuration of an inactive tenant. When a request for an inactive tenant with an incompatible configuration arrives, impact is contained to a single worker. The configuration load architecture now prefers serving as many customers as quickly as possible, rather than waiting until everything is ready before serving anyone. The above changes are slated to complete in April 2026 and will bring our RTO from the current ~1 hour to under 10 minutes – for complete recovery from a worst case scenario. Continuous validation through Game Days A critical element of our recovery confidence comes from GameDay fault-injection testing. We don’t simply design recovery mechanisms and assume they work—we break the system deliberately and observe how it responds. Since late 2025, we have conducted recurring GameDay drills that simulate the exact failure scenarios we are defending against: Food Taster crash scenarios: Injecting deliberately faulty tenant configurations, to verify that they are caught and isolated with zero impact on live traffic. In our January 2026 GameDay, the Food Taster process crashed as expected, the system halted the update within approximately 5 seconds, and no customer traffic was affected. Master process crash scenarios: Triggering master process crashes across test environments to verify that workers continue serving traffic, that the Local Config Shield engages within 10 seconds, and that the coordinated recovery tool restores full operation within the expected timeframe. Multi-region failure drills: Simulating simultaneous failures across multiple regions to validate that global Config Shield mechanisms engage correctly, and that recovery procedures scale without requiring manual per-region intervention. Fallback test drills for critical Azure services running behind Azure Front Door: In our February 2026 GameDay, we simulated the complete unavailability of Azure Front Door, and successfully validated failover for critical Azure services with no impact to traffic. These drills have both surfaced corner cases and built operational confidence. They have transformed recovery from a theoretical plan into tested, repeatable muscle memory. As we noted in an internal communication to our team: “Game day testing is a deliberate shift from assuming resilience to actively proving it—turning reliability into an observed and repeatable outcome.” Closing Part 1 of this series emphasized preventing unsafe configurations from reaching the data plane, and data plane resiliency in case an incompatible configuration reaches production. This post has shown that prevention alone is not enough—when failures do occur, recovery must be fast, predictable, and bounded. By ensuring that the FlatBuffer cache is never invalidated, by loading only active tenants, and by building safe coordinated recovery tooling, we have transformed failure handling from a fleet-wide crisis into a controlled operation. These recovery investments work in concert with the prevention mechanisms described in Part 1. Together, they ensure that the path from incident detection to full service restoration is measured in minutes, with customer traffic protected at every step. In the next post of this series, we will cover the third pillar of our resiliency strategy: tenant isolation—how micro-cellular architecture and ingress-layered sharding can reduce the blast radius of any failure to a small subset, ensuring that one customer’s configuration or traffic anomaly never becomes everyone’s problem. We deeply value our customers’ trust in Azure Front Door. We are committed to transparently sharing our progress on these resiliency investments, and to exceed expectations for safety, reliability, and operational readiness.Azure Front Door: Implementing lessons learned following October outages
Abhishek Tiwari, Vice President of Engineering, Azure Networking Amit Srivastava, Principal PM Manager, Azure Networking Varun Chawla, Partner Director of Engineering Link to Part 2 - Azure Front Door Resiliency Series Link to Part 3 - Azure Front Door Resiliency Series Introduction Azure Front Door is Microsoft's advanced edge delivery platform encompassing Content Delivery Network (CDN), global security and traffic distribution into a single unified offering. By using Microsoft's extensive global edge network, Azure Front Door ensures efficient content delivery and advanced security through 210+ global and local points of presence (PoPs) strategically positioned closely to both end users and applications. As the central global entry point from the internet onto customer applications, we power mission critical customer applications as well as many of Microsoft’s internal services. We have a highly distributed resilient architecture, which protects against failures at the server, rack, site and even at the regional level. This resiliency is achieved by the use of our intelligent traffic management layer which monitors failures and load balances traffic at server, rack or edge sites level within the primary ring, supplemented by a secondary-fallback ring which accepts traffic in case of primary traffic overflow or broad regional failures. We also deploy a traffic shield as a terminal safety net to ensure that in the event of a managed or unmanaged edge site going offline, end user traffic continues to flow to the next available edge site. Like any large-scale CDN, we deploy each customer configuration across a globally distributed edge fleet, densely shared with thousands of other tenants. While this architecture enables global scale, it carries the risk that certain incompatible configurations, if not contained, can propagate broadly and quickly which can result in a large blast radius of impact. Here we describe how the two recent service incidents impacting Azure Front Door have reinforced the need to accelerate ongoing investments in hardening our resiliency, and tenant isolation strategy to mitigate likelihood and the scale of impact from this class of risk. October incidents: recap and key learnings Azure Front Door experienced two service incidents; on October 9 th and October 29 th , both with customer-impacting service degradation. On October 9 th : A manual cleanup of stuck tenant metadata bypassed our configuration protection layer, allowing incompatible metadata to propagate beyond our canary edge sites. This metadata was created on October 7 th , from a control-plane defect triggered by a customer configuration change. While the protection system initially blocked the propagation, the manual override operation bypassed our safeguards. This incompatible configuration reached the next stage and activated a latent data-plane defect in a subset of edge sites, causing availability impact primarily across Europe (~6%) and Africa (~16%). You can learn more about this issue in detail at https://aka.ms/AIR/QNBQ-5W8 On October 29 th : A different sequence of configuration changes across two control-plane versions produced incompatible metadata. Because the failure mode in the data-plane was asynchronous, the health checks validations embedded in our protection systems were all passed during the rollout. The incompatible customer configuration metadata successfully propagated globally through a staged rollout and also updated the “last known good” (LKG) snapshot. Following this global rollout, the asynchronous process in data-plane exposed another defect which caused crashes. This impacted connectivity and DNS resolutions for all applications onboarded to our platform. Extended recovery time amplified impact on customer applications and Microsoft services. You can learn more about this issue in detail at https://aka.ms/AIR/YKYN-BWZ We took away a number of clear and actionable lessons from these incidents, which are applicable not just to our service, but to any multi-tenant, high-density, globally distributed system. Configuration resiliency – Valid configuration updates should propagate safely, consistently, and predictably across our global edge, while ensuring that incompatible or erroneous configuration never propagate beyond canary environments. Data plane resiliency - Additionally, configuration processing in the data plane must not cause availability impact to any customer. Tenant isolation – Traditional isolation techniques such as hardware partitioning and virtualization are impractical at edge sites. This requires innovative sharding techniques to ensure single tenant-level isolation – a must-have to reduce potential blast radius. Accelerated and automated recovery time objective (RTO) – System should be able to automatically revert to last known good configuration in an acceptable RTO. In case of a service like Azure Front Door, we deem ~10 mins to be a practical RTO for our hundreds of thousands of customers at every edge site. Post outage, given the severity of impact which allowed an incompatible configuration to propagate globally, we made the difficult decision to temporarily block configuration changes in order to expedite rollout of additional safeguards. Between October 29 th to November 5 th , we prioritized and deployed immediate hardening steps before opening up the configuration change. We are confident that the system is stable, and we are continuing to invest in additional safeguards to further strengthen the platform's resiliency. Learning category Goal Repairs Status Safe customer configuration deployment Incompatible configuration never propagates beyond Canary · Control plane and data plane defect fixes · Forced synchronous configuration processing · Additional stages with extended bake time · Early detection of crash state Completed Data plane resiliency Configuration processing cannot impact data plane availability Manage data-plane lifecycle to prevent outages caused by configuration-processing defects. Completed Isolated work-process in every data plane server to process and load the configuration. January 2026 100% Azure Front Door resiliency posture for Microsoft internal services Microsoft operates an isolated, independent Active/Active fleet with automatic failover for critical Azure services Phase 1: Onboarded critical services batch impacted on Oct 29 th outage running on a day old configuration Completed Phase 2: Automation & hardening of operations, auto-failover and self-management of Azure Front Door onboarding for additional services March 2026 Recovery improvements Data plane crash recovery in under 10 minutes Data plane boot-up time optimized via local cache (~1 hour) Completed Accelerate recovery time < 10 minutes March 2026 Tenant isolation No configuration or traffic regression can impact other tenants Micro cellular Azure Front Door with ingress layered shards June 2026 This blog is the first in a multi-part series on Azure Front Door resiliency. In this blog, we will focus on configuration resiliency—how we are making the configuration pipeline safer and more robust. Subsequent blogs will cover tenant isolation and recovery improvements. How our configuration propagation works Azure Front Door configuration changes can be broadly classified into three distinct categories. Service code & data – these include all aspects of Azure Front Door service like management plane, control plane, data plane, configuration propagation system. Azure Front Door follows a safe deployment practice (SDP) process to roll out newer versions of management, control or data plane over a period of approximately 2-3 weeks. This ensures that any regression in software does not have a global impact. However, latent bugs that escape pre-validation and SDP rollout can remain undetected until a specific combination of customer traffic patterns or configuration changes trigger the issue. Web Application Firewall (WAF) & L7 DDoS platform data – These datasets are used by Azure Front Door to deliver security and load-balancing capabilities. Examples include GeoIP data, malicious attack signatures, and IP reputation signatures. Updates to these datasets occur daily through multiple SDP stages with an extended bake time of over 12 hours to minimize the risk of global impact during rollout. This dataset is shared across all customers and the platform, and it is validated immediately since it does not depend on variations in customer traffic or configuration steps. Customer configuration data – Examples of these are any customer configuration change—whether a routing rule update, backend pool modification, WAF rule change, or security policy change. Due to the nature of these changes, it is expected across the edge delivery / CDN industry to propagate these changes globally in 5-10 mins. Both outages stemmed from issues within this category. All configuration changes, including customer configuration data, are processed through a multi-stage pipeline designed to ensure correctness before global rollout across Azure Front Door’s 200+ edge locations. At a high level, Azure Front Door’s configuration propagation system has two distinct components - Control plane – Accepts customer API/portal changes (create/update/delete for profiles, routes, WAF policies, origins, etc.) and translates them into internal configuration metadata which the data plane can understand. Data plane – Globally distributed edge servers that terminate client traffic, apply routing/WAF logic, and proxy to origins using the configuration produced by the control plane. Between these two halves sits a multi-stage configuration rollout pipeline with a dedicated protection system (known as ConfigShield): Changes flow through multiple stages (pre-canary, canary, expanding waves to production) rather than going global at once. Each stage is health-gated: the data plane must remain within strict error and latency thresholds before proceeding. Each stage’s health check also rechecks previous stage’s health for any regressions. A successfully completed rollout updates a last known good (LKG) snapshot used for automated rollback. Historically, rollout targeted global completion in roughly 5–10 minutes, in line with industry standards. Customer configuration processing in Azure Front Door data plane stack Customer configuration changes in Azure Front Door traverse multiple layers—from the control plane through the deployment system—before being converted into FlatBuffers at each Azure Front Door node. These FlatBuffers are then loaded by the Azure Front Door data plane stack, which runs as Kubernetes pods on every node. FlatBuffer Composition: Each FlatBuffer references several sub-resources such as WAF and Rules Engine schematic files, SSL certificate objects, and URL signing secrets. Data plane architecture: o Master process: Accepts configuration changes (memory-mapped files with references) and manages the lifecycle of worker processes. o Workers: L7 proxy processes that serve customer traffic using the applied configuration. Processing flow for each configuration update: Load and apply in master: The transformed configuration is loaded and applied in the master process. Cleanup of unused references occurs synchronously except for certain categories à October 9 outage occurred during this step due to a crash triggered by incompatible metadata. Apply to workers: Configuration is applied to all worker processes without memory overhead (FlatBuffers are memory-mapped). Serve traffic: Workers start consuming new FlatBuffers for new requests; in-flight requests continue using old buffers. Old buffers are queued for cleanup post-completion. Feedback to deployment service: Positive feedback signals readiness for rollout.Cleanup: FlatBuffers are freed asynchronously by the master process after all workers load updates à October 29 outage occurred during this step due to a latent bug in reference counting logic. The October incidents showed we needed to strengthen key aspects of configuration validation, propagation safeguards, and runtime behavior. During the Azure Front Door incident on October 9 th , that protection system worked as intended but was later bypassed by our engineering team during a manual cleanup operation. During this Azure Front Door incident on October 29 th , the incompatible customer configuration metadata progressed through the protection system, before the delayed asynchronous processing task resulted in the crash. Configuration propagation safeguards Based on learnings from the incidents, we are implementing a comprehensive set of configuration resiliency improvements. These changes aim to guarantee that any sequence of configuration changes cannot trigger instability in the data plane, and to ensure quicker recovery in the event of anomalies. Strengthening configuration generation safety This improvement pivots on a ‘shift-left’ strategy where we want to ensure that we catch regression early before they propagate to production. It also includes fixing the latent defects which were the proximate cause of the outage. Fixing outage specific defects - We have fixed the control-plane defects that could generate incompatible tenant metadata under specific operation sequences. We have also remediated the associated data-plane defects. Stronger cross-version validation - We are expanding our test and validation suite to account for changes across multiple control plane build versions. This is expected to be fully completed by February 2026. Fuzz testing - Automated fuzzing and testing of metadata generation contract between the control plane and the data plane. This allows us to generate an expanded set of invalid/unexpected configuration combinations which might not be achievable by traditional test cases alone. This is expected to be fully completed by February 2026. Preventing incompatible configurations from being propagated This segment of the resiliency strategy strives to ensure that a potentially dangerous configuration change never propagates beyond canary stage. Protection system is “always-on” - Enhancements to operational procedures and tooling prevent bypass in all scenarios (including internal cleanup/maintenance), and any cleanup must flow through the same guarded stages and health checks as standard configuration changes. This is completed. Making rollout behavior more predictable and conservative - Configuration processing in the data plane is now fully synchronous. Every data plane issue due to incompatible meta data can be detected withing 10 seconds at every stage. This is completed. Enhancement to deployment pipeline - Additional stages during roll-out and extended bake time between stages serve as an additional safeguard during configuration propagation. This is completed. Recovery tool improvements now make it easier to revert to any previous version of LKG with a single click. This is completed. These changes significantly improve system safety. Post-outage we have increased the configuration propagation time to approximately 45 minutes. We are working towards reducing configuration propagation time closer to pre-incident levels once additional safeguards covered in the Data plane resiliency section below are completed by mid-January, 2026. Data plane resiliency The data plane recovery was the toughest part of recovery efforts during the October incidents. We must ensure fast recovery as well as resilience to configuration processing related issues for the data plane. To address this, we implemented changes that decouple the data plane from incompatible configuration changes. With these enhancements, the data plane continues operating on the last known good configuration—even if the configuration pipeline safeguards fail to protect as intended. Decoupling data plane from configuration changes Each server’s data plane consists of a master process which accepts configuration changes and manages lifecycle of multiple worker processes which serve customer traffic. One of the critical reasons for the prolonged outage in October was that due to latent defects in the data plane, when presented with a bad configuration the master process crashed. The master is a critical command-and-control process and when it crashes it takes down the entire data plane, in that node. Recovery of the master process involves reloading hundreds of thousands of configurations from scratch and took approximately 4.5 hours. We have since made changes to the system to ensure that even in the event of the master process crash due to any reason - including incompatible configuration data being presented - the workers remain healthy and able to serve traffic. During such an event, the workers would not be able to accept new configuration changes but will continue to serve customer traffic using the last known good configuration. This work is completed. Introducing Food Taster: strengthening config propagation resiliency In our efforts to further strengthen Azure Front Door’s configuration propagation system, we are introducing an additional configuration safeguard known internally as Food Taster which protects the master and worker processes from any configuration change related incidents, thereby ensuring data plane resiliency. The principle is simple: every data-plane server will have a redundant and isolated process – the Food Taster – whose only job is to ingest and process new configuration metadata first and then pass validated configuration changes to active data plane. This redundant worker does not accept any customer traffic. All configuration processing in this Food Taster is fully synchronous. That means we do all parsing, validation, and any expensive or risky work up front, and we do not move on until the Food Taster has either proven the configuration is safe or rejected it. Only when the Food Taster successfully loads the configuration and returns “Config OK” does the master process proceed to load the same config and then instruct the worker processes to do the same. If anything goes wrong in the Food Taster, the failure is contained to that isolated worker; the master and traffic-serving workers never see that invalid configuration. We expect this safeguard to reach production globally in January 2026 timeframe. Introduction of this component will also allow us to return closer to pre-incident level of configuration propagation while ensuring data plane safety. Closing This is the first in a series of planned blogs on Azure Front Door resiliency enhancements. We are continuously improving platform safety and reliability and will transparently share updates through this series. Upcoming posts will cover advancements in tenant isolation and improvements to recovery time objectives (RTO). We deeply value our customers’ trust in Azure Front Door. The October incidents reinforced how critical configuration resiliency is, and we are committed to exceeding industry expectations for safety, reliability, and transparency. By hardening our configuration pipeline, strengthening safety gates, and reinforcing isolation boundaries, we’re making Azure Front Door even more resilient so your applications can be too.
