PostgreSQL trigram GIST vs. GIN indexes

Question

I have a PostgreSQL 9.1 database that contains 10M+ rows and some text fields that need similarity and %word%-like searches, so I decided to use trigram indexes.

Initially, I started to use GIN indexes, but now I am wondering whether I should rather use GIST.

The theory says:

• GIN index lookups are about three times faster than GiST
• GIN indexes take about three times longer to build than GiST
• GIN indexes are moderately slower to update than GiST indexes, but about 10 times slower if fast-update support was disabled (see Section 58.4.1 for details)
• GIN indexes are two-to-three times larger than GiST indexes

I have tried to create both GIN and GIST indexes and found the following in practice (on my dataset):

• For LIKE '%word%' queries, GIN is about the same speed (or even 5-20% slower) as GIST for the first time (when the index for the trigrams in the query are not cached yet).
• For LIKE '%word%' queries, GIN is about 5 times faster than GIST if the trigrams in the query were searched recently. GIST has always the same speed, regardless of cachedness.
• For % 'word' (similarity) queries, GIN is about 5-8 times slower than GIST, depending on the cachedness of the index.
• GIN is about 10% smaller than GIST. However, it seems to grow faster if there are UPDATE statements to the table. Unless, of course, I VACUUM FULL often enough.

So I see some discrepancies between theory and practice:

1. speed: GIST is often faster, except GIN is totally in memory (Can you ensure that in advance?), and even then only for LIKE '%word%' queries.
2. size: GIN is actually smaller. Is this typical for trigram indexes or is my case an outlier?
3. May the behaviour be significantly different in 9.4?
4. A last question: is it wise to use GIN if there are about 10-50k UPDATE statements and 5-10k INSERT statements daily, and weekly VACUUMis probably reasonable (not much usage on the weekend), or will it really hurt UPDATE and INSERT performance?