Managed instances operations

Operations guides for managed instances.

Red/black deployment model

At any point in time only one deployment is the active production instance, this is noted in deploy-sourcegraph-managed/$CUSTOMER/terraform.tfvars, and except during upgrades only one is running. You can see this via:

$ gcloud compute instances list --project=sourcegraph-managed-$CUSTOMER
default-red-instance  us-central1-f  n1-standard-8                       RUNNING

The NAME value indicates the currently active instance (red or black). During an upgrade, both default-red-instance and default-black-instance will be present. One being production, the other being the “new” upgraded production instance. When the upgrade is complete, the old one is turned off (red/black swap). Learn more about managed instances upgrades here.

Accessing the instance

For CSE’s, also refer to Accessing Managed Instances.

SSH access

Locate the GCP instance you’d like to access (usually either default-red-instance or default-black-instance), and then:

$ gcloud beta compute ssh --zone "us-central1-f" --tunnel-through-iap --project "sourcegraph-managed-$CUSTOMER" default-$DEPLOYMENT-instance

If you get an error:

ERROR: (gcloud.beta.compute.start-iap-tunnel) Error while connecting [4003: u'failed to connect to backend'].
ssh_exchange_identification: Connection closed by remote host
ERROR: (gcloud.beta.compute.ssh) [/usr/bin/ssh] exited with return code [255].

This may be indicating that the VM is currently not running—check:

$ gcloud beta compute instances list --project=sourcegraph-managed-$CUSTOMER

And start the instance if needed (e.g. through the web UI.)

Accessing the Docker containers

SSH into the relevant instance and then:

sudo su
docker ps

You can then use regular Docker commands (e.g. docker exec -it $CONTAINER sh) to interact with the containers.

Restarting for configuration updates

  1. SSH into the relevant instance
  2. cd to the /deployment/docker-compose folder and run:
  3. docker compose restart sourcegraph-frontend-0 sourcegraph-frontend-internal


Locate the GCP instance you’d like to access (usually either default-red-instance or default-black-instance), and then:

export PORT=80 # or one of the below ports
gcloud compute start-iap-tunnel default-$DEPLOYMENT-instance $PORT --local-host-port=localhost:4444 --zone "us-central1-f" --project "sourcegraph-managed-$CUSTOMER"

This will port-forward localhost:4444 to port 80 on the VM instance. Some common ports:

Note that other ports are prevented by the allow-iap-tcp-ingress firewall rule.

Access through the GCP load balancer as a user would

Users access Sourcegraph through GCP Load Balancer/HTTPS -> the Caddy load balancer/HTTP -> the actual sourcegraph-frontend/HTTP. If you suspect that an issue is being introduced by the GCP load balancer itself, e.g. perhaps a request is timing out there due to some misconfiguration, then you will need to access through the GCP load balancer itself. If the managed instance is protected by the load balancer firewall / not publicly accessible on the internet, then it is not possible for you to access $ as a normal user would.

You can workaround this by proxying your internet traffic through the instance itself - which is allowed to reach and go through the public internet -> the GCP load balancer -> back to the instance itself. To do this, create a SOCKS5 proxy tunnel over SSH (replace default-red-instance, if needed):

bash -c '(gcloud beta compute ssh --zone "us-central1-f" --tunnel-through-iap --project "sourcegraph-managed-$CUSTOMER" default-$DEPLOYMENT-instance -- -N -p 22 -D localhost:5000) & sleep 600; kill $!'

Then test you can access it using curl:

$ curl --proxy socks5://localhost:5000 https://$
<a href="/sign-in?returnTo=%2F">Found</a>.

You can now reproduce the request using curl, or configure your OS or browser to use socks5://localhost:5000:

  • Windows: Follow the “using the SOCKS proxy” section in this article [mirror] to enable it on Internet Explorer, Edge and Firefox.
  • macOS: System Preferences → Network → Advanced → Proxies → check “SOCKS proxy” and enter the host and the port.
  • Linux: Most browsers have proxy settings in their Settings/Preferences.
  • Command-line apps: Many CLIs accept http_proxy or https_proxy environment variables or arguments you can set the proxy. Consult the help or the manpage of the program.

IMPORTANT: Once you are finished, terminate the original gcloud beta compute ssh command so your machine’s traffic is no longer going over the instance. The command above will automatically terminate after 10 minutes, to prevent this.

Finding the external IPs

$ gcloud compute addresses list --project=sourcegraph-managed-$CUSTOMER
default-global-address   $GLOBAL_IP      EXTERNAL                                         IN_USE
default-nat-manual-ip-0  $NAT_IP_ONE     EXTERNAL                    us-central1          IN_USE
default-nat-manual-ip-1  $NAT_IP_TWO     EXTERNAL                    us-central1          IN_USE
  • $GLOBAL_IP is the IP address that $ should point to, it is the address of the GCP Load Balancer.
  • $NAT_IP_ONE and $NAT_IP_TWO are the external IPs from which egress traffic from the deployment will originate from. These are the addresses from which Sourcegraph will access the customer’s code host, and as such the customer will need to allow them access to e.g. their internal code host.

Resizing Disks

Disk storage can be safely increased on managed instances at any time. Do not try to decrease the disk size - Terraform will destroy and recreate the disk causing data loss.

To increase the disk size:

  1. Set up your variables as usual.

  2. Increase the value of data_disk_size in terraform.tfvars and run terraform apply

  3. Commit and push your changes:

    git add terraform.tfvars && git commit -m "$CUSTOMER: increase disk size"
    git push origin HEAD
  4. Follow the GCP instructions to resize the block storage. In most cases, the commands will look like:

    ../util/ "sudo resize2fs /dev/sdb"

    Then confirm the new size is visible:

    ../util/ "df -h /dev/sdb"

Running these commands will have no impact on a running deployment and can be safely performed without interruption to the customer.

Capturing network traffic for analysis on the instance

In some cases, you may need to capture network traffic for debugging issues on the instance. We use Wireshark and tcpdump to do this. First, find the service you are interested in, for example, to capture traffic to/from the sourcegraph-frontend service:

  sudo tcpdump -i any -s 65535 'port 3080' -w /tmp/sourcegraph-frontend.pcap

Next you need to scp this from the instance:

   # after eval $(mg workon)
   gcloud compute scp root@default-$DEPLOYMENT-instance:/tmp/sourcegraph-frontend.pcap . # copy from instance

Open the pcap file in Wireshark (installable with brew install --cask wireshark)

Changing the instance


The state of managed instances infrastructure and deployment artifact are stored in the following repositories

We are aligned with the company-wide testing philosophy. All changes to above repositories have to be done via a Pull Request, and the Pull Request requires a test plan in the description to detail how to validate the change. Additionally, the Pull Request will require at least one approval prior to merging. This ensure we establish a proper audit trail of what’s changed and the reason behind it.

Availability of the instance

Uptime Checks


We are aligned with the company-wide incident response playbook to handle managed instances downtime.

We utilize GCP Uptime Checks to perform uptime checks against the managed instance frontend url. When such alert is fired, it usually means the service is completely not accessible to customers. In the event of downtime, GCP will notify On-Call DevOps engineers via Opsgenie and the On-Call engineers will proceed with our incident playbook to ensure we reach to a resolution.

Performance Checks


We utilize the Sourcegraph built-in alerting to monitor application-level metrics. We identify a list of critical metrics that are representation on the overall system performance, and the alert is delivered to Opsgenie. Opsgenie will notify On-Call DevOps engineers](../ and the On-Call engineers will proceed to investigate and ensure we reach to a resolution.

A list of critial metrics that will be routed to Opsgenie:

Confirm instance health


The primary tool that monitors releases post-deployment are through a variety of uptime monitors and system performance metrics. These metrics are covered in documentation related to SOC/CI-87.

Following a release upgrade, in addition to automated instance health checks, we will perform additional manul check to confirm instance health.

Run command below and inspect the output to ensure that all containers are healthy (in particular, look for anything that says Restarting):

mg --customer $CUSTOMER check

Access Grafana and confirm the instance is healthy by verifying no critical alerts are firing, and there has been no large increase in warning alerts:

mg forward grafana

Check frontend logs and there are no recent errors

mg ssh-exec docker logs sourcegraph-frontend-0

Instance technicalities

Impact of recreating the instance via Terraform

All configuration about the instance itself is stored in Terraform, so recreating the instance is a non-destructive action. A brand new VM will be provisioned by Terraform, the startup script will initialize it and mount the existing data disk back into the VM, and the Sourcegraph Docker containers will start up.

This typically involves around 8m40s of downtime: 6m destroying the network endpoint group, and 2m creating the new instance / installing software / launching Docker services.

Instance is recreated when startup script changes

Any time a startup script is changed, Terraform will plan to delete the old VM instance and recreate it such that the script runs again.

This is a non-destructive action, aside from the fact that it involves downtime for the deployment.

Debugging startup scripts

View startup script logs

cat /var/log/syslog | grep startup-script

Run startup script and debug:

sudo google_metadata_script_runner --script-type startup --debug

WARNING: Running our startup script twice is a potentially harmful action, as it is usually only ran once.

More details:

Viewing container logs

Containers logs are persisted in GCP Logging by utilizing the GCP Logging Driver.

Let’s say you want to check the logs of sourcegraph-frontend-0.

Visit and ensure you’re in the right GCP project. Then you may write the following query:

There’s a Show query toggle at the top right corner, turn it on

log_name="projects/sourcegraph-managed-dev/logs/gcplogs-docker-driver" : sourcegraph-frontend-0

Learn more about the query language syntax

Fix corrupted repo on gitserver

Context of why this exists:

A broken repo can be identified by

Once you have identified a repo is constantly failing to be updated/fetched, execute the following steps:

  1. Set up env vars

    export PROJECT_PREFIX=sourcegraph-managed
    export DEPLOYMENT=$(gcloud compute instances list --project "$PROJECT_PREFIX-$CUSTOMER" | grep -v "executors" | awk 'NR>1 { if ($1 ~ "-red-") print "red"; else print "black"; }')
    export CUSTOMER=<customer_or_instance_name>
  2. Determine if git prune or git fetch is failing by exec’ing into the gitserver-0 container

mg ssh
docker exec -it gitserver-0 sh
cd /data/repos/<repo_name>
git prune && git fetch
# look for errors, no output indicates clean repo
  1. Run the following script to have repo-updater queue an update


Disaster Recovery and Business Continuity Plan


Follow restore process


FAQ: “googleapi: Error 400: The network_endpoint_group resource … is already being used”

If terraform apply is giving you:

Error: Error when reading or editing NetworkEndpointGroup: googleapi: Error 400: The network_endpoint_group resource 'projects/sourcegraph-managed-$COMPANY/zones/us-central1-f/networkEndpointGroups/default-neg' is already being used by 'projects/sourcegraph-managed-$COMPANY/global/backendServices/default-backend-service', resourceInUseByAnotherResource

Or similar—this indicates a bug in Terraform where GCP requires an associated resource to be deleted first and Terraform is trying to delete (or create) that resource in the wrong order.

To workaround the issue, locate the resource in GCP yourself and delete it manually and then terraform apply again.