- We need to persist all client session in database.
- This involve persisting client session’s id and configuration in mysql server. persisting entire session object is not required. We need to do this as first thing while creating user session.
- For each request from client lens server should not validate session id from its in memory map. It should rather validate it from persisted session in mysql server.
session is present in mysql and not present in lens server’s in memory map(session manager) then it should create a lens session with same session id and add it to in memory map
session is not present in mysql and present in server’s in memory map then it signifies that user have initiated close session from other host and we will need to remove this session from current host’s in memory map as well
session is not present in mysql and in server’s in memory map then return “invalid session”.
- Whenever session expired/closed then we need to remove this session from lens server’s in memory map and mysql.
- For sync queries lens server will close connection as soon as it is stopped or failed with appropriate failure message. Client will be retrying the query.
- For async-light queries and async-heavy queries lens server takes care of rescheduling all queries which were in running or queued state at the time of restart. we can use this to make query ids available with both servers.
- For async queries we need to persist all non finished async user queries in mysql server in new table current_query.
- If user executes query on host1 and it goes down and nginx points to host2 then user should be able to poll on query status. Lens server will first check query id in its inmemory maps if it is not present then it will check in “finished_query” table in mysql else in “current_query” table in mysql. In this case we will continue showing old status of query since it is scheduled by host1. As soon as host1 comes up it will reschedule these queries and will change the status. Optionally we can have host2 to move these queries in “allqueries” map of its query service which will take care of recovering, in this case we don't need to wait for host1 to come up and reschedule.
- While moving query from ”finished_quey” map to “finished_query” table in mysql we will need to remove it from “current_query” tables as well.
- This flow will remain same for request ids/query ids generated for query plan/estimate.
- Lens downtime will be zero in case of deployment and if primary fail.
- User will not have to create a new session whenever switch happens
- User will be able to get status of all async queries seamlessly.