ERRINJ_VY_READ_PAGE and ERRINJ_VY_READ_PAGE_TIMEOUT injections are
used by vinyl/errinj test to check that a page read error is handled
properly. The cache checks if these injections are enabled and bails
out if so. Since commit a31c2c10 ("vinyl: force read view in
iterator in autocommit mode"), this is not necessary, because cache
is not used unless SELECT is called from a transaction, and the above
mentioned test doesn't use transactions. So let's remove the checks.
If we ever enable cache for all SELECTs, we can disable cache in the
test with box.cfg.vinyl_cache instead of using error injections.

Currently, new entries are added to the cache even if the limit is set
to 0, which is obviously incorrect. Fix it. This might be useful for
debugging and/or performance evaluation.

Closes #3172

Make box.cfg.vinyl_cache option dynamically configured. If the new
value is less than the current cache usage, it will block until it
reclaims enough cache entries to fit in the new limit.

If the user terminates tarantool while compaction is in progress,
tarantool won't exit immediately - it will hang until all dump and
compaction tasks that are currently in progress are complete, which
may take quite a while on large data sets. What is especially funny,
once a task has been finished, vinyl will not commit the produced run
file in the vylog, because the event loop has already been stopped,
and so will delete it after restart and schedule the task anew.

This patch makes the scheduler forcefully abort all running tasks
as soon as possible.

Closes #3166

This is pointless as worker threads don't use coio.

vy_run_writer incapsulates logic of writing statements to a run.
It provides API to write statements one by one. It is needed so
that we can abort a run writing task before waiting for it to
finish writing the file.

Edited by @locker:
 - add region_truncate() wherever necessary
 - do not reallocate row index ibuf for each page
 - remove vy_run_write(), use vy_run_writer directly instead

Needed for #3166

If a replica receives some changes then corresponding wal events are
generated and relay fiber is waken up before heartbeat timeout.
But there may be nothing to send if all changes are from the current
relay peer. In this case an applier doesn't receive anything and break a
connection.

Fixes #3160

Add memory occupied by cached slabs to the size reported by
arena_size and arena_used_ratio respectively.

This should make memory monitoring easier:

if items_used_ratio > 0.9 and arena_used_ratio > 0.9 and quota_used_ratio > 0.9 then
    we run out of memory
end

On FreeBSD we have either system iconv (part of libc) or the one
in /usr/local. The first one is not fully compatible with Linux/
OSX iconv and both adds specific prefixes to function names.
So to use them via Lua FFI mechanism, specific names are required.
Proposal fix:
Under FreeBSD link with /usr/local/lib/libiconv.so and add prefix.

Closes gh-3073.

* Add parser of cookies from 'Set-Cookie' header.
* All cookies are stored in response object in cookie table.
* The format of each table element is following:
resp[key] = {value, {option, ...}},
where key, value is cookie key, value
options are array of strings, set-cookie header options,
e.g. HttpOnly, Expires=<date>.

Closes #2801

* delete old small parser with nginx tested one
* functionality is not changed

If a key is updated after a secondary index scan, but before a
primary index lookup, then ignore this update.

Closes #2442

If a column is nullable and is the last defined one (via index parts
or space format), it can be omited on insertion. Such absent fields
are treated as NULLs in comparators and are not stored.

Closes #2988

This template parameter serves to detect, if two sequences of
key parts are actually parts of tuples. In such a case
key_compare_parts does additional comparison by primary key parts,
if a NULL was met.

This parameters is actually a crutch, put by me to avoid code
doubling in tuple_compare_sequential_nullable. And this parameter
complicates #2988, where a sequential nullable tuple contains
optional parts.

Needed for #2988

Closes #2048

They have no sense without key_def.h.

Needed for #2048

Update comments and move space_def_check_compatibility to
space_def.[hc].

On a single field two restrictions exist: tuple format and index part.

The proposal is to allow to use different types for different restrictions,
if these types are compatible. And for tuples validation use the most
strict type.

For example, one field can be number in space format, unsigned in one index,
integer in another index and scalar in a third index. Configuration like this
can be allowed, because all of these types can store unsigned. And unsigned is
used to validate tuples.

Besides, it allows incremental DDL. For example, now if you want to change a
type of a field, which is used by several indexes, then it is impossible on
a non-empty space, even if a new type is less strict than the old one.
Lets allow this too.

Closes #3008

At first, now if a space is not empty, its format can not be
restricted by nullability removal, or by restriction of a field type.

The workaround was format removal and reset. Lets allow normal format
restriction followed by space format checking.

At second, now an index is rebuild on any change of its key definition.
But there is no sense to rebuild it, if uniqueness is turned off, part
field numbers and collations are not changed, because it means, that
only types was restricted or extended. So format checking is enough.

Needed for #3008

Compatibility must be commutative, but this function is not
commutative. It checks, that one type can store values of another
type, but not conversely.

Closes #2973

Not all workloads need bloom filters enabled for all indexes. Let's
allow to disable them on per-index basis by setting bloom_fpr to 1.
This will allow to save some memory if bloom filters are unused.

Closes #3138

Currently, one can set insane values for most vinyl index options, which
will most certainly result in a crash (e.g. bloom_fpr = 100). Add some
sanity checks.

While a node of the cluster is re-bootstrapping (joining again),
other nodes may try to re-subscribe to it. They will fail, because
the rebootstrapped node hasn't tried to subscribe hence hasn't been
added to the _cluster table yet and so is not present in the hash
at the subscriber's side for replica_on_applier_reconnect() to look
it up.

Fix this by making a subscriber create an id-less (REPLICA_ID_NIL)
struct replica in this case and reattach the applier to it. It will
be assigned an id when it finally subscribes and is registered in
_cluster.

Fixes 71b33405 replication: reconnect applier on master rebootstrap

If box.cfg() successfully connects to a number of replicas sufficient to
form a quorum (>= box.cfg.replication_connect_quorum), it won't return
until it syncs with all of them (lag <= box.cfg.replication_sync_lag).
If one of the replicas forming a quorum disconnects permanently while
sync is in progress, box.cfg() will hang forever.

Such a behavior is rather unreasonable. After all, syncing a quorum is
best-effort. It would be much more sensible to return from box.cfg()
leaving the instance in the 'orphan' mode in this case. This patch does
exactly that: now if we detect that not enough replicas are connected to
form a quorum while we are syncing we stop syncing immediately.

Currently, the max time box.cfg() may wait for connection to replicas to
be established is hardcoded to box.cfg.replication_timeout times 4. As
a result, users can't revert to pre replication_connect_quorum behavior,
when box.cfg() blocks until it connects to all replicas. To fix that,
let's introduce a new configuration option, replication_connect_timeout,
which determines the replication configuration timeout. By default the
option is set to 4 seconds.

Closes #3151

If a unique index includes all parts of another unique index, we can
skip the check for duplicates for it on INSERT. Let's mark all such
indexes with a special flag on CREATE/ALTER and optimize out the check
if the flag is set. If there are two indexes that index the same set
of fields, check uniqueness for the one with a lower id, because it
is likelier to have have a "warmer" cache.

Closes #3154

We keep run files corresponding to (at least) the last snapshot, because
we need them for backups and replication. Deletion of compacted run
files is postponed until the next snapshot. As a consequence, we don't
delete run files created on a replica during the join stage. However, in
contrast to run files created during normal operation, these are pure
garbage and should be deleted right away. Not deleting them can result
in depletion of disk space, because vinyl has quite high write
amplification by design.

We can't write a functional test for this, because there's no way to
guarantee that compaction started during join will finish before join
completion - if it doesn't, compacted runs won't be removed, because
they will be assigned to the snapshot created by join.

Closes #3162

Vinyl index key definition is stored in vylog even if an index is
empty, and we while do not have a method to update it. So vinyl
index key definition alter is forbidden even on an empty space.

Closes #3169

Closes #3173

If any ddl operation is in progress then all other ddls are
waiting on schema latch. But after first ddl will be done any
other request may be issued just after it and commit order will be
broken in case of multimaster replication.  To prevent this
behavior any ddl operation should wait until all queued ddls are
done.

Fixes #2951

Prevent latch lock interception by other already scheduled or active
fiber if there is only one waiting. This is needed for strict latch
ordering.