When I actually run a node, have it do a few snapshots, and then restart it -- the CLI is getting "not a database".

Don't merge this, not until I understand the errors (and if they are real, why the test suite isn't catching them).

Seems fine, can't reproduce the issue via the CLI. I've also added end-to-end testing to ensure restoring to both in-mem and on-disk databases works fine.

I got the system into this state again.

127.0.0.1:4001> select count(*) from foo
ERR! file is not a database
127.0.0.1:4001> select count(*) from foo
ERR! file is not a database
127.0.0.1:4001> 
EOF (CTRL+D)
~/repos/rqlite/src/github.com/rqlite/rqlite/cmd/rqlite (no-disk-during-restore)$ ./rqlite 
Welcome to the rqlite CLI. Enter ".help" for usage hints.
Connected to version 5
127.0.0.1:4001> select count(*) from foo
ERR! file is not a database
127.0.0.1:4001> select count(*) from foo
ERR! file is not a database
127.0.0.1:4001> select count(*) from foo
ERR! file is not a database
127.0.0.1:4001> 

Not good.

The test case seems to be:

• start a node using in-mem database
• trigger snapshot
• restart node
• queries break, with "not a database".

Very strange that none of the current testing catches this, only the command line testing.

I think I see the issue. This problem doesn't occur 100% of the time with the following script, but pretty close:

#!/bin/bash
DATA_DIR=`mktemp -d`

$RQLITED -raft-snap 13 -raft-snap-int=1s $DATA_DIR &
sleep 3

$RQBENCH -b 10 -m 1 -n 500 -o 'CREATE TABLE foo (id INTEGER NOT NULL PRIMARY KEY, name TEXT)' 'INSERT INTO foo(name) VALUES("fiona")'

killall rqlited
sleep 3

$RQLITED -raft-snap 50000 -raft-snap-int=500s $DATA_DIR

SQLite is using the memory passed in, and making changes to it -- that memory represents the database after all. But that memory is in the Go memory space. And I'm sure this is allowed, since Go is a garbage collected language.