You are obviously using a fake-IMMUTABLE function wrapper for unaccent() to use it in indexes. It's based on the search dictionary "unaccent" implicitly or explicitly. The dictionary to be used can be provided as optional first function parameter to unaccent(). The name of the dictionary can be schema-qualified. Else it needs to reside in one of the schemas of the current search_path.
I do not advise to manipulate the function unaccent() provided by the extension. Instead, create and use a wrapper function like instructed here:
In your case, the additional module unaccent might be installed in a schema that's missing from the search_path of the session reporting these errors. When you try manually, you do that in a session where the search_path includes that same schema, so no error there.
You can either set your search_path to include the schema of the extension in all sessions, or schema-qualify the dictionary in the function calls of unaccent().
Security for client programs has been tightened with Postgres 10.3 / 9.6.8 etc. Now you need to schema-qualify function and dictionary when used in any indexes - as demonstrated in the linked answer above.
The release notes for Postgres 10.3:
Also, the changes described in the second changelog entry below may
cause functions used in index expressions or materialized views to
fail during auto-analyze, or when reloading from a dump. After
upgrading, monitor the server logs for such problems, and fix affected
functions.
...
Avoid use of insecure search_path settings in pg_dump and other client programs (Noah Misch, Tom Lane)
pg_dump, pg_upgrade, vacuumdb and other PostgreSQL-provided
applications were themselves vulnerable to the type of hijacking
described in the previous changelog entry; since these applications
are commonly run by superusers, they present particularly attractive
targets. To make them secure whether or not the installation as a
whole has been secured, modify them to include only the pg_catalog
schema in their search_path settings. Autovacuum worker processes now
do the same, as well.
In cases where user-provided functions are indirectly executed by
these programs — for example, user-provided functions in index
expressions — the tighter search_path may result in errors, which will
need to be corrected by adjusting those user-provided functions to not
assume anything about what search path they are invoked under. That
has always been good practice, but now it will be necessary for
correct behavior.
Bold emphasis mine.
See the special mention of auto-analyze and autovacuum? To fix, schema-qualify function name and dictionary in the unaccent() call of the wrapper function used in your index(es) as demonstrated in the linked answer. Like:
CREATE OR REPLACE FUNCTION f_unaccent(text)
RETURNS text AS
$func$
SELECT public.unaccent('public.unaccent', $1) -- schema-qualify function and dictionary
$func$ LANGUAGE sql IMMUTABLE;
You can change an existing function definition with CREATE OR REPLACE FUNCTION ... without rebuilding existing indexes based on it. Make sure the function keeps using the same function & dictionary, then indexes keep working properly without REINDEX. (You can't DROP the function, though, while it's used in an index.)
Moving the unaccent module to the system catalog pg_catalog also works. Then everything is in the search_path automatically. You ended up doing so, but I do not advocate that. Messing with system catalogs is generally discouraged and can easily damage your installation if you don't know exactly what you are doing.
Related:
The search dictionary "unaccent" is also regularly used by text search configurations in full text search. Related:
search_pathfor autovacuum has been changed for security reasons. I don't expect them to revert that change nor to offer users the unsafe option to do so. Have you tried to schema-qualify function and dictionary as suggested?CREATE OR REPLACE FUNCTION ...without rebuilding anything. Expression indexes rely on the immutability of the used function and are not stopping you. If you make sure that the function keeps using the same function & dictionary, indexes keep working properly. (You can'tDROPthe function, though.)