There is not guarantee or requirement for WAL writers to only add
series entries in increasing order of IDs. A postings list cannot look
back and thus unordered WAL entries would skip over IDs to not truncate
from the WAL.
We replace it with a simple boolean check function that does not require
order.
IDs for new series are handed out before the postings are locked. Thus
series are not indexed in order of their IDs, which could result in only
partially sorted postings list.
Iterating over those silently skipped elements as the sort invariant was
violated.
We were not properly maintaining the scrape cache when the same metric
was exposed with a different string representation.
This overall reduces the scraping cache's complexity, which fixes the
issue and saves about 10% of memory in a scraping-only Prometheus
instance.
The problem reported in #2799 was that in the event that all records for a
name were removed, the target group was never updated to be the "empty" set.
Essentially, whatever Prometheus last saw as a non-empty list of targets
would stay that way forever (or at least until Prometheus restarted...). This
came about because of a fairly naive interpretation of what a valid-looking
DNS response actually looked like -- essentially, the only valid DNS responses
were ones that had a non-empty record list. That's fine as long as your
config always lists only target names which have non-empty record sets; if
your environment happens to legitimately have empty record sets sometimes,
all hell breaks loose (otherwise-cleanly shutdown systems trigger up==0 alerts,
for instance).
This patch is a refactoring of the DNS lookup behaviour that maintains
existing behaviour with regard to search paths, but correctly handles empty
and non-existent record sets.
RFC1034 s4.3.1 says there's three ways a recursive DNS server can respond:
1. Here is your answer (possibly an empty answer, because of the way DNS
considers all records for a name, regardless of type, when deciding
whether the name exists).
2. There is no spoon (the name you asked for definitely does not exist).
3. I am a teapot (something has gone terribly wrong).
Situations 1 and 2 are fine and dandy; whatever the answer is (empty or
otherwise) is the list of targets. If something has gone wrong, then we
shouldn't go updating the target list because we don't really *know* what
the target list should be.
Multiple DNS servers to query is a straightforward augmentation; if you get
an error, then try the next server in the list, until you get an answer or
run out servers to ask. Only if *all* the servers return errors should you
return an error to the calling code.
Where things get complicated is the search path. In order to be able to
confidently say, "this name does not exist anywhere, you can remove all the
targets for this name because it's definitely GORN", at least one server for
*all* the possible names need to return either successful-but-empty
responses, or NXDOMAIN. If any name errors out, then -- since that one
might have been the one where the records came from -- you need to say
"maintain the status quo until we get a known-good response".
It is possible, though unlikely, that a poorly-configured DNS setup (say,
one which had a domain in its search path for which all configured recursive
resolvers respond with REFUSED) could result in the same "stuck" records
problem we're solving here, but the DNS configuration should be fixed in
that case, and there's nothing we can do in Prometheus itself to fix the
problem.
I've tested this patch on a local scratch instance in all the various ways I
can think of:
1. Adding records (targets get scraped)
2. Adding records of a different type
3. Remove records of the requested type, leaving other type records intact
(targets don't get scraped)
4. Remove all records for the name (targets don't get scraped)
5. Shutdown the resolver (targets still get scraped)
There's no automated test suite additions, because there isn't a test suite
for DNS discovery, and I was stretching my Go skills to the limit to make
this happen; mock objects are beyond me.
The Client type is already exposed, but can't be used without the config for it
also being exposed. Using the remote.Client from other programs is useful to do
full end-to-end tests of Prometheus's remote protocol against adapter
implementations.
This can happen in the situation where the system scales up the number of shards massively (to deal with some backlog), then scales it down again as the number of samples sent during the time period is less than the number received.
