1521 delete tables on upgrade

[absoludity]

Ref: #1521

Following on from #1526, this PR adds a pre-upgrade hook which invalidates the database cache of charts. When postgres is being used, this drops and recreates the tables. For mongodb, it's currently a no-op. I've not added functionality to check for a schema change at the moment, so this will rebuild the cache on every upgrade.

IRL test:
Initial deploy dev environment and verify schema is the original (ie. without namespace field in repos table):

$ k -n kubeapps exec kubeapps-postgresql-0 -- sh -c 'PGPASSWORD=$POSTGRES_PASSWORD psql -d assets -U $POSTGRES_USER -c "\d repos;"'
                                    Table "public.repos"
   Column    |       Type        | Collation | Nullable |              Default
-------------+-------------------+-----------+----------+-----------------------------------
 id          | integer           |           | not null | nextval('repos_id_seq'::regclass)
 name        | character varying |           |          |
 checksum    | character varying |           |          |
 last_update | character varying |           |          |
Indexes:
    "repos_pkey" PRIMARY KEY, btree (id)
    "repos_name_key" UNIQUE CONSTRAINT, btree (name)

Upgrade with the new sync image (which has the new schema):

helm -n kubeapps upgrade --reuse-values kubeapps ./chart/kubeapps/ --set apprepository.syncImage.tag=2 --set apprepository.image.tag=1

See preupgrade job deployed successfully (before I added the hook-delete-policy):

$ k -n kubeapps logs kubeapps-internal-apprepository-jobs-preupgrade-jt9zz
time="2020-02-21T00:58:32Z" level=info msg="Successfully invalidated cache"

Re-check the table and now see it's updated.

  $ k -n kubeapps exec kubeapps-postgresql-0 -- sh -c 'PGPASSWORD=$POSTGRES_PASSWORD psql -d assets -U $POSTGRES_USER -c "\d repos;"'
                                    Table "public.repos"
   Column    |       Type        | Collation | Nullable |              Default
-------------+-------------------+-----------+----------+-----------------------------------
 id          | integer           |           | not null | nextval('repos_id_seq'::regclass)
 namespace   | character varying |           | not null |
 name        | character varying |           | not null |
 checksum    | character varying |           |          |
 last_update | character varying |           |          |
Indexes:
    "repos_pkey" PRIMARY KEY, btree (id)
    "repos_namespace_name_key" UNIQUE CONSTRAINT, btree (namespace, name)
Referenced by:
    TABLE "charts" CONSTRAINT "charts_repository_id_fkey" FOREIGN KEY (repository_id) REFERENCES repos(id)

Verified cache is being repopulated as app repo starts.

[absoludity]

I'm assuming this is an old job which no longer exists, as I couldn't see the name being used anywhere.

[andresmgot]

LGTM

[andresmgot]

is the CASCADE necessary?

[absoludity]

Not here it's not, thanks. I'd originally only deleted the one table and relied on the cascades, but I had to change that as the files table doesn't have a FK in this PR. I've removed the CASCADE and will re-add it when the FK's are in the next PR (I think)>

[andresmgot]

this is not really necessary since it will be executed with syncing but I am okay if you want to leave it to leave the tables ready

[absoludity]

Yeah, the reason for having it here is so that it is sync'd with the correct code (the new, rather than from the still running older pods). We will probably need a transaction here to avoid race conditions with the existing pods too.