Skip to content

Nightshift and your cloud

A pod that crash-loops with password authentication failed is not a Kubernetes problem. Neither is a 502 behind a Front Door whose WAF blocks the request, or a pod that cannot reach its database because a private DNS zone lost its link to the cluster’s network. When the evidence of an investigation points at a cloud resource like these, Nightshift reads that resource’s settings and says what is wrong and how to fix it.

Nightshift reads a resource’s settings and state: its firewall rules, its private endpoint, its power state, a secret’s versions and dates. It never reads a secret value, the contents of a blob, a key or an app setting. A wrong secret is found from its version and date and from the dependency refusing it, never from the value. See what Seylo reads in your cloud.

Every cloud read is made by Seylo’s server through your keyless cloud account, never by the agent in your cluster, and nothing is written to your cloud. Every fix in your cloud is a command or code change a person runs.

From what it already reads in the cluster, never from a Secret’s value:

Where What it looks for
Environment variables and ConfigMaps the workload uses Hosts such as *.postgres.database.azure.com, *.mysql.database.azure.com, *.database.windows.net, *.vault.azure.net, *.blob.core.windows.net, *.azurewebsites.net, *.azurefd.net, a private IP address
SecretProviderClass (Secrets Store CSI driver) The Key Vault’s name and the objects it mounts - names only
The ServiceAccount The Azure workload identity it uses (azure.workload.identity/client-id)
Ingresses Hosts and addresses a Front Door origin should point at

A host that appears only inside a Secret is not seen, and the investigation says so, for example env DB_HOST (container api) comes from Secret orders-db; Seylo does not read Secret values, so a host in it is not seen. An architecture file can name such a dependency.

  • Only during an investigation, after its opening reads. Never on a schedule.
  • Only when the evidence points at a resource it found. A log line or an event must name its host, or be the only resource of a kind that fits the error (a bare timeout that names no host is not enough). Otherwise nothing is read in your cloud. A database’s timeout expired counts as a failing connection, like connection refused.
  • Again, once, when a later step names a host. The first read follows the opening reads. If a later step, such as the pod’s logs, names a cloud host the first read did not cover, Nightshift reads your cloud once more, while it has no cloud finding yet.
  • At most 3 resources per investigation.
Resource What Nightshift can find
Key Vault Public access off with no private endpoint; a firewall that does not allow the cluster; a private DNS zone that is missing, not linked to the cluster’s network or pointing at the wrong address; a secret that does not exist, is disabled or has expired; a SecretProviderClass pinned to an old version; pods still holding an old version after the secret rotated; an identity with no role or access policy to read secrets; a pod template missing the workload identity label, or an identity with no federated credential for the cluster
Storage account The same network and DNS causes; an identity with no data role on the account; shared key access turned off while the app uses the account key
PostgreSQL and MySQL flexible servers, Azure SQL The server is not ready (stopped, for example); its firewall does not admit the address the pods leave from (see below); its storage is full; it rejects the password the pods hold; it ran out of connection slots
Virtual machines and the network path No network interface holds the address the app calls; the VM is stopped or deallocated; a peering that is missing or not connected; a route that drops the traffic or sends it to an appliance that is gone or stopped; an NSG rule that denies it
App Service and Function Apps The app is stopped, named as such (Function App fn-orders is Stopped); its access restrictions refuse the address the pods leave from, read in priority order with the final “Deny all”; the app itself returns server errors
Front Door An origin that points at none of the cluster’s ingress hosts, or at the wrong ports; a failing health probe; a WAF policy in Prevention mode blocking the requests; an origin host header no Ingress serves; an HTTPS-only route to a host with no TLS on the Ingress

Nightshift evaluates NSG rules and routes itself, from the subnet’s and the network interface’s NSGs, Azure’s default rules, the route table and the peerings. Routes learned over BGP from a VPN or ExpressRoute gateway are not visible to it, and it says so rather than calling a route missing.

Not yet diagnosed: Cosmos DB, Azure Cache for Redis and Service Bus.

A firewall or an access restriction is only half the answer; the other half is the address your pods come from. Nightshift compares the two and says, for example: orders-db's firewall allows only 203.0.113.10; the pods leave from 198.51.100.7 (the AKS cluster's outbound public IP), which is not admitted. The fix is the az command that adds that exact address.

The address is the one the database or app reported, when it reported one. Otherwise Nightshift reads where the AKS cluster leaves from, by its outbound type:

Outbound type Where Nightshift reads the address
Load balancer The load balancer’s outbound public IPs
Managed NAT gateway The NAT gateway’s outbound public IPs
User-assigned NAT gateway The NAT gateway on the node subnet, and its public IPs
User-defined routing (all traffic through a firewall) The node subnet’s route table, its 0.0.0.0/0 route to an appliance, the Azure Firewall at that address, and that firewall’s public IPs

When the address cannot be known, the finding says not known and why, and the command carries <the cluster's outbound IP> for you to fill in. Seylo never guesses an address. This happens when:

  • the next hop is a third-party appliance, not an Azure Firewall: its public address is not stated anywhere Seylo reads;
  • there is no 0.0.0.0/0 route, or it does not go to an appliance;
  • the NAT gateway uses a public IP prefix, which is not expanded;
  • a rule is a service tag or a subnet, which Seylo does not expand.

Azure’s “Allow Azure services” rule (0.0.0.0) admits an AKS cluster’s public addresses, so it is never blamed.

Each finding appears under Cloud dependencies in the investigation panel:

Part What it says
Confidence and root cause One sentence, for example Private DNS zone privatelink.vaultcore.azure.net is not linked to the cluster's VNet aks-vnet, so the pods resolve kv-orders.vault.azure.net to its public address, which kv-orders refuses.
Evidence Numbered, like every other piece of evidence: the log lines that showed the failure, and each cloud fact read. Hover a number to read it
Fix in the cloud The exact az command, and Show Terraform / Bicep for the same change as code. A person runs it, or it goes into your infrastructure repository as a pull request. Seylo makes no change in your cloud
Fix in the cluster When part of the fix is inside the cluster, such as a missing label or a restart, it is proposed through Nightshift’s normal approval, like any other fix
A person’s step What only a person can do, such as setting a new secret value, or choosing a connection limit. Seylo does not pick a value it has not measured
Verified by What shows the fix worked, for example the 403s stopping
Not seen What Seylo could not read, and why, with the one thing that would let it

When nothing is found, the panel says so: The cloud resources the evidence names were read; nothing in their configuration explains the failure. When the evidence names a resource in a subscription that is not connected: The evidence names a cloud resource, but Seylo cannot read it.

When pods mount a secret through the Secrets Store CSI driver and the vault’s current version is newer than the one the pods hold - rotation is off, or the app read the value once at start - the fix is a rolling restart. It waits for a person, or runs on its own under a standing grant. Nightshift reads which version each pod mounted from the CSI driver’s SecretProviderClassPodStatus objects; if the agent’s role may not read them, it says so, and the restart is not proposed.

Application Insights and Front Door logs: failures only

Section titled “Application Insights and Front Door logs: failures only”

When an App Service or Function App the evidence names has an Application Insights component connected, or a Front Door profile sends its diagnostic logs to a workspace, Nightshift runs one query, scoped to that resource, never to a whole workspace:

  • Failures only: failed requests, exceptions, failed dependency calls; for Front Door, health probes that did not succeed.
  • Counts, not rows: grouped by type, target and result code, with at most one sample message per exception type, cut to 200 characters and redacted.
  • Never custom dimensions, URL query strings, request or response bodies, or user and session ids.
  • Bounded: from 10 minutes before the failure was first seen, never more than 60 minutes back, at most 30 rows.

With no Application Insights component connected to the app, the investigation says so and gives the az monitor app-insights commands that add one.

Some reads are billed in your own cloud account, not by Seylo. Nightshift reads only during an investigation, at most 3 resources, so these are small; but they are yours. Prices are as each provider states them, read on 2026-11-09. Where a provider states no price, the row says so.

Cloud Read Price as stated Free allowance Source
Azure Log Analytics query over Analytics tables Free: “Querying Analytics Logs is free of charge” - Azure Monitor Logs cost
Azure Log Analytics query touching Basic or Auxiliary tables $0.005 per GB scanned (East US), $0.00615 per GB (Central US), meter “Search Queries Scanned” None listed Azure Retail Prices API; Azure Monitor Logs cost
Azure Metrics through the batch API $0.01 per 1,000 calls First 1M calls a month Azure Retail Prices API
Azure Metrics through Azure Resource Manager Unpriced, because Microsoft’s own sources conflict on whether it is billed - -
Azure Resource Graph Free (“As a free service”) - Resource Graph overview
Azure Resource reads through Azure Resource Manager, Key Vault secret properties included Unpriced, because Microsoft states throttling limits for them, not a price. No Key Vault data-plane call is made - -
AWS CloudWatch GetMetricData $0.01 per 1,000 metrics requested (US East, N. Virginia) None: always charged AWS Price List API; CloudWatch pricing
AWS CloudWatch Logs Insights query $0.005 per GB scanned (US East) 5 GB, shared with ingestion and storage AWS Price List API
AWS Secrets Manager DescribeSecret, ListSecrets Unpriced, because AWS does not say whether these count as API calls - Secrets Manager pricing
AWS CloudTrail event history (90 days) No charge stated for management event history - CloudTrail pricing
Google Cloud Cloud Logging reads No additional charge for querying log data - Google Cloud Observability pricing
Google Cloud Cloud Monitoring timeSeries.list $0.50 per million time series returned First 1M time series per billing account Google Cloud Observability pricing
Google Cloud Secret Manager metadata Free (“Management operations”) - Secret Manager pricing

The AWS and Google Cloud rows are what those reads cost when they are made. In this release Nightshift’s diagnosis reads Azure only. For your bill read, see prices from your cloud bill.