Date
Authors
Andy Nash (Unlicensed) Joseph (Pepe) Kelly John Simmons (Deactivated) Andy Dingley Simon Meredith (Unlicensed) Paul Hoang (Unlicensed)
Status
Date |
|
Authors | Andy Nash (Unlicensed) Joseph (Pepe) Kelly John Simmons (Deactivated) Andy Dingley Simon Meredith (Unlicensed) Paul Hoang (Unlicensed) Sachin Mehta (Unlicensed) |
Status | LiveDefect resolved. Actions being ticketed up (Andy Nash (Unlicensed)) |
Summary | On Friday evening we noticed the server used to build our applications and run scheduled jobs was struggling. It then crashed, subsequently causing ETL (scheduled data flows between systems) and data related issues. Systems used to search, transfer and display data in TIS along with data stores then also froze between Friday and Saturday |
Impact | Users unable to search for trainees. Staging environment was not updated with production data on Saturday as scheduled No data synchronisation between systems |
Timeline
0600 Friday 16th October 2020 | We have a system to automatically upgrade versions of software, and keep these versions and dependencies between systems synchronised. This ensures software is patched and up to date. A large number of these upgrades were triggered at one time which caused our build server to build the application multiple times in order to run the automated tests against them prior to deployment. |
0630 Friday 16th October 2020 | One of our search servers failed |
0730 Friday 16th October 2020 | More search server failures Stage environment suffers various server failures |
1630 Friday 16th October | Build server fails |
0110 Saturday 17th October | High number of messages in ESR system waiting to be processed |
0600 Saturday 17th October | ESR scheduled job process failed to start |
0800 Saturday 17th October | Multiple server failures by this point and TIS search non-operational |
Root Causes
An update to the operating system was automatically applied without any need to restart servers
This caused a conflict of versions with our containerisation platform that we use to host our applications. This is the first time we have seen this since TIS went live so we need to assess if this is a one-off occurrence
The trainee data used for our search function was unavailable (we use elasticsearch but this was unreachable)
The version conflict caused a domino effect which resulted in multiple server failures. The was compounded by our automated dependency system (Dependabot) which created multiple versions of the application which caused the build server to fail (not enough RAM)
Many downstream processes rely on the build server
We haven’t configured the monthly Dependabot sweep to stagger the automated builds
Trigger
6am patch caused a version conflict which required a restart of the platform running our applications (Docker)
Failure of Docker to restart put pressure on the build server as well as many other servers
ElasticSearch issues compromised the search function for users of the app
Dependencies of many timed jobs on the build server being available
Resolution
Restarted build server
Restarted National Data Warehouse ETLs on Stage and Prod environments
Restarted Docker on search server
Restarted Neo4J GraphDB
Turned on Change Data Capture system to deal with backlog of data changes from TIS to ESR
Removed all the old images from MongoDB, one of our databases
Detection
Good news is that there were lots of monitoring alerts in various channels in Slack and from the graphs, below, there was probably enough evidence there to have alerted us to a problem needing resolution
Users pointed out the issue with search not functioning correctly on TIS
Bad news is we didn’t act on the initial Friday indicators (they didn’t appear to be as serious an indicator as they were)
Actions
Look into being able to do a full restart automatically if a similar situation arises
Increase the memory (RAM) on the build server
Implement and close any remaining patches to ESR which were automatically created
Restrict the number of dependencies that can be upgraded at one time to reduce the load on the build server
Make changes to build server to move from concurrent to sequential builds
Look at removing the nightly job that refreshes search data as this may no longer be necessary
Lessons Learned (Good and Bad)
Dev team to check the monitoring channels more regularly
Act on anything unusual (resolve immediately if easy to do so, alert others immediately if it appears serious, raise on stand ups otherwise).
Incident has encouraged us to do the most thorough route cause analysis we have ever done, and subsequently identify some inefficiencies in related areas, and to map out a range of actions short and longer term to address them all.
...
Technical Detail
Summary
On Friday evening we saw Jenkins struggling, and then fell over, subsequently causing ETL and data related issues
...
Impact
...
No Stage. No Prod. No data syncing in various places
...
...
...
...
...
...
. Elasticsearch, RabbitMQ and MongoDB then also fell over between Friday and Saturday
Timeline (collation and interpretation from across all Slack monitoring channels
...
)
First trigger | |||
| Friday |
| |
...
...
|
...
|
...
|
...
|
...
|
...
|
...
|
...
|
...
| |||
Starting Friday and continuing over the weekend | |||
| |||
...
| Friday |
|
...
|
...
|
...
...
|
...
|
...
|
...
| because of sharding, users may have been seeing only partial results at this point) (PH) 👈 need to look into and confirm this assumption… | ||
Then | |||
...
|
...
|
...
|
...
|
...
|
...
|
...
|
...
|
...
| Saturday |
|
...
|
...
2020-10-17 07:43 Prod Mongo goes down (PH)
…at this point ☝ pretty much everything is being effected by the combination of issues
...
2020-10-17 08:48 
(see Prometheus graph below)
...
2020-10-17 09:27 (see Prometheus graph below)
...
2020-10-17 10:23 D/B Prod/Stage sync started but never completed
...
2020-10-17 10:23 NDW ETL: Stage (PaaS) failed
...
2020-10-18 02:42 ESR Sentry errors x 7 (reappearance of the same issue across all services)
...
2020-10-18 07:37 TCS ES sync job failed to run/complete on either blue or green servers
...
2020-10-18 10:23 NDW ETL: Stage (PaaS) failed
| |||
…at this point ☝ pretty much everything is being effected by the combination of issues | |||
👇 shall I remove this section of the timeline completely from this incident log? Seems overkill, given pretty much everything was down at this point | |||
| |||
...
|
...
|
...
|
...
|
...
|
...
|
...
|
...
|
...
Root Cause(s)
Dependabot does a monthly check on tooling versions and auto-generates PRs and builds which, in the case of the ESR area, creates many many concurrent containers which trips Jenkins up (not enough RAM)
Lots and lots of downstream processes rely on Jenkins being up
We haven’t been able to configure the monthly Dependabot sweep to avoid a massive concurrent hit of PRs / builds
Trigger
Configuration of Dependabot inadvertently puts strain on Jenkins
Dependencies of many timed jobs on Jenkins being available
Not enough configuration of retries
Resolution
Manually brought Jenkins back up
Manually restarted everything that had been affected by it being down
Detection
Good news is that there were lots of monitoring alerts in various channels in Slack and from the graphs, below, there was probably enough evidence there to have alerted us to a problem needing resolution
Bad news is we didn’t act on the initial Friday indicators, and John was rebuilding his machine over the weekend, so couldn’t do his normal knight in shiny armour stuff!
Action Items
...
Confirmed action Items
...
Owner
...
Status
...
Comments
...
0. Spike the creation an Ansible config to do a full restart (John Simmons (Deactivated) to complete the description of this) - in case we come across this situation happening again.
Can we put a checklist of actions together that were taken to try to get us up and running again?
Ideally this checklist could be a ref for future failures
...
Ops
...
To do
...
1. bump up the Jenkins RAM to 32Gb (short term ONLY). Add a reminder to revisit this in 1 month / 2 months?
...
Ops
...
Done
...
2. disable integration tests on ESR projects for PR pipeline (they’d still run on merge to master, rather than each PR). These are what fire up the local stack and test containers (hold back on…'if we’re not planning to do any further ESR work once Leeerrroooy leave, we could just disable the integration tests in ESR')
ESR guys not keen at all on this. Risk of breaking the pipeline and blocking anything thereafter. When you need them the most… don’t have tests switched off!
2. Code in AWS S3 creds to save firing up local stack. Need to be careful setting this up. Shared bucket for integration tests across Reval, TISSS as well. With archiving rules deleting everything >5 days old.
See ‘diet’ ticket below - lots of work, but an investment that will have its uses across all products within the TIS portfolio
...
ESR & others
...
To do
...
3. Close outstanding ESR major version change PRs - how many is ‘critical mass’? But without being blazé about approving PRs. Ignore minor versions.
...
Sachin
...
To do
...
Consider ways to reduce system resources instead?
Discussed with Doris and ESR team last week…
...
4. Restrict the number of PRs Dependabot opens on each ESR project to 1 and to major versions only (given they’re microservices, we still might get significant numbers).
There are dependencies between PRs. So restricting to one might have a chain reaction. No getting round that things like this will need manual intervention.
So let’s see how restricting to major versions affects things and decide on further action if needed at that point.
...
AndyD
...
To do
...
AndyD / Pepe
...
To do
...
Pipelines not written to take advantage of extra nodes - is the time investment required greater than the benefit
...
6. Address https://hee-tis.atlassian.net/browse/TISNEW-5613 quickly (removing local stack from the Data exporter process).
Needs refining and sub-tasking with ESR and perm team, before ESR disappear.
...
ESR & perm team
...
To do
...
Spike ticket completed this Sprint. This ticket is the result of that Spike.
...
7. The ElasticSearch nightly sync shouldn’t be necessary. Verify that ElasticSearch is being updated properly during the day.
...
Pepe
...
To do
...
TBC
Lessons Learned (Good and Bad)
Check monitoring channels in Slack, check Prometheus, check Grafana as a matter of professional pride, daily.
Act* on anything unusual (* resolve yourself if you can, alert others immediately if you think it’s serious, raise on stand ups otherwise).
Incident has encouraged us to identify the route cause, and to identify some inefficiencies in related areas, and to map out a range of actions short and longer term to address them all.
Techy Stuff
5 whys
...
Everything fell over
...
Comments following catch up 2020-10-21
...
Jenkins
ESR containers taking up all the resource
Too many (Dependabot) PRs outstanding, builds, rebasing
ESR did not had time to action them because of the launch of new world code
...
Underlying OS upgrade occurred and was applied (LivePatch)
Probably something to do with that very specific update (it’s never happened before and has been running since the Apps were built
Did Docker struggle with the patch, rather than Amazon making a mistake with applying the patch (no complaints on Amazon forums, so looks quite specific to us)
Version of Docker is linked to whatever’s available in
aptNot worth changing our set up for the sake of a ‘freak’ occurrence that probably will only happen once every 3 years? Or switching off the OS upgrades and handling it manually every week / switch off everything and restart it regularly?
...
Initial discussion, along with short and longer term actions
What can we do about Dependabot creating and building simultaneously?
Dependabot does run sequentially, but much faster than Jenkins can process things so everything appears concurrent.
We could get Dependabot to add a GitHub label to the PR - add something to the Jenkins file to read the label and mark as “Don’t run”. But this stops Dependabot being useful.
ESR preoccupied with launching New World, understandably!
Can perm team keep on top of ESR stuff when they leave?
Even when keeping on top of things, will it eventually be too much anyway?
Or is it simply a case of the team not controlling the overall number of open PRs?
Original Jenkins build was never designed to handle this much load - underlying architecture isn’t there for the level of automation we now have. It is designed for a single node, not load-balancing
Is Jenkins the right tool for everything it’s being asked to do? No:
bump up the Jenkins RAM to 32Gb (short term ONLY). Add a reminder to revisit this in 1 month / 2 months?
disable integration tests on ESR projects for PR pipeline (they’d still run on merge to master, rather than each PR). These are what fire up the local stack and test containers (hold back on…'if we’re not planning to do any further ESR work once Leeerrroooy leave, we could just disable the integration tests in ESR')
Close outstanding ESR PRs - how many is ‘critical mass’? But without being blazé about approving PRs
Restrict the number of PRs Dependabot opens on each ESR project to 1 (but given they’re microservices, it still might be a big number). Not much of a concern if we do 2. above.
The ElasticSearch nightly sync shouldn’t be necessary. Verify that ElasticSearch is being updated properly during the day.
move ETLs over to ECS tasks (serverless ‘run container' instructions to AWS - not reliant on our infrastructure).
This would remove the dependency on Jenkins - so if it went down, the jobs could continue.
Don’t do scheduled jobs / anything with a timer - use Cron server instead for this stuff.
Just use Jenkins as a build server (Metabase also runs on Jenkins, but doesn’t use much)ticket up addressing our infrastructure so that the set up ESR have created does run - it’s been done right!
get ourselves a dedicated Jenkins server (what size
)move to ElasticSearch SaaS
Tech notes
Related to TISNEW-5454
ESR Data Exporter changed - tried to build all containers - tried to fire up all the outstanding Dependabot (monthly) PRs (each PR spun up 30 odd containers).
Jenkins box only has 16Gb of RAM - can we bung up to 32? But this doubles the cost. Can we hit the route cause, rather than the size of RAM.
Turn off dependabot for now? Automatic rebasing? Not keen to switch off a valuable resource. Better to address root cause.
Modify jobs to only build if, e.g.:
PR was raised by dependabot (and is a major version change)
There are other PRs for the project building
Single Jenkins node responsibility for control and building everything.
We haven’t moved to GHA yet.
Person Sync (rather than Person Owner?) Job:
Prod-to-Stage didn’t failed on Saturday
First a problem with Elastic Search Sync job on Sunday
Docker restart made elastic-search available to ADMINS-UI again
Checklist of actions we took in this instance
- Restarted Jenkins
- Restarted NDW ETLs on Stage (current and PaaS)
- Restarted Docker on ES nodes
- Restarted Neo4J container (rather than the service)
- Spun up CDC (Paul) - looking at the HUGE backlog of messages (which started coming down rapidly. Phew!)
- Removed all the dangling volumes from Mongo (several times) (old images that are taking up space)
| Saturday |
|
Root Cause(s)
OS
aptpatch running every morning used the LivePatch function to apply the patch without needing to restart everything.This caused conflict with Docker - Note, however, that this conflict has not been seen before since TIS’s inception. So the working theory is that this was a one-off conflict that is not likely to reoccur.
The index of trainees for the searchpage (elasticsearch) was unreachable and couldn’t start up.
The conflict then compromised everything else in a domino effect, compounded by the backed up Dependabot PRs and builds which, in the case of the ESR area, creates multiple concurrent containers which trips Jenkins up (not enough RAM)
Many downstream processes rely on Jenkins being up
We haven’t configured the monthly Dependabot sweep to stagger the hit of PRs / builds
Trigger
6am OS patch process auto-applied the patches via LivePatch, causing a conflict with Docker which needed to be restarted
Failure of Docker restarting following the OS patch puts strain on ElasticSearch and Jenkins (and everything else)
ElasticSearch issues compromised the search function for users of the app
Dependencies of many timed jobs on Jenkins being available
Not enough configuration of retries
Resolution
Restarted Jenkins
Restarted NDW ETLs on Stage (current and PaaS)
Restarted Docker on ES nodes
Restarted Neo4J container (rather than the service)
Spun up CDC (Paul) - looking at the HUGE backlog of messages from changes to TIS data (which started coming down rapidly. Phew!)
Removed all the dangling volumes from Mongo (several times) (old images that are taking up space)
Detection
Good news is that there were lots of monitoring alerts in various channels in Slack and from the graphs, below, there was probably enough evidence there to have alerted us to a problem needing resolution
Users pointed out the issue with search not functioning correctly on TIS
Bad news is we didn’t act on the initial Friday indicators (they didn’t appear to be as serious an indicator as they were)
Action Items
Confirmed action Items | Owner | Status | Comments |
|---|---|---|---|
0. Spike the creation an Ansible config to do a full restart (John Simmons (Deactivated) to complete the description of this) - in case we come across this situation happening again. | Ops | To do | |
1. bump up the Jenkins RAM to 32Gb (short term ONLY). Add a reminder to revisit this in 1 month / 2 months? | Ops | Done | |
| ESR & others | To do | |
3. Close outstanding ESR major version change PRs - how many is ‘critical mass’? But without being blazé about approving PRs. Ignore minor versions. | Sachin | To do | Consider ways to reduce system resources instead? |
4. Restrict the number of PRs Dependabot opens on each ESR project to 1 and to major versions only (given they’re microservices, we still might get significant numbers). | AndyD | To do | |
5. Use Jenkins pipeline rather than node - to make things sequential rather than concurrent | AndyD / Pepe | To do | Pipelines not written to take advantage of extra nodes - is the time investment required greater than the benefit |
6. Address https://hee-tis.atlassian.net/browse/TISNEW-5613 quickly (removing local stack from the Data exporter process). | ESR & perm team | To do | Spike ticket completed this Sprint. This ticket is the result of that Spike. |
7. The ElasticSearch nightly sync shouldn’t be necessary. Verify that ElasticSearch is being updated properly during the day. | Pepe | To do | TBC |
Lessons Learned (Good and Bad)
Check monitoring channels in Slack, check Prometheus, check Grafana as a matter of professional pride, daily.
Act* on anything unusual (* resolve yourself if you can, alert others immediately if you think it’s serious, raise on stand ups otherwise).
Incident has encouraged us to do the most thorough route cause analysis we have ever done, and subsequently identify some inefficiencies in related areas, and to map out a range of actions short and longer term to address them all.
...
Techy Stuff
5 whys
Everything fell over | Comments following catch up 2020-10-21 | |
|---|---|---|
|
| Jenkins could to with some TLC. ES went down too (before Jenkins). However, there is an underlying OS update issue that triggered everything. |
Initial discussion, along with short and longer term actionsWhat can we do about Dependabot creating and building simultaneously? Dependabot does run sequentially, but much faster than Jenkins can process things so everything appears concurrent.
ESR preoccupied with launching New World, understandably! Can perm team keep on top of ESR stuff when they leave? Even when keeping on top of things, will it eventually be too much anyway? Original Jenkins build was never designed to handle this much load - underlying architecture isn’t there for the level of automation we now have. It is designed for a single node, not load-balancing Is Jenkins the right tool for everything it’s being asked to do? No:
| ||
Checklist of actions should take in future
One-off task: Look at the way updates are applied - AWS system manager - tell it when you want to do the patching. Schedule to patch Stage first, check and then apply to Prod / Then set up a simple Slack reminder to check all is well
- Manage user expectations - alert them AS SOON AS we know something’s amiss. AND keep updates flowing as we resolve things. UNDERSTAND what of our remedial actions will have what affect on users. UPDATE status alert on TIS (ensure it’s not just Phil KOTN who has access / does this
- When an ES node fails, generate an auto-restart
- Check logs
- Platform-wide restart - ensure everything is brought back up ‘clean’
- Recheck
Graphs
...