Refuse to boot if we don't have at least 1 GiB per shard, unless in developer
mode.
The primary violator here is docker, but since it starts in developer mode,
it won't get fixed. We need some extra logic for this case.
Message-Id: <20161221090222.28677-1-avi@scylladb.com>
"This patchset contains fixes for the changes introduced in "Query result
size limiting". It also improves handling of short data reads.
I order to minimise chances of digest mismatch during data queries replicas
that were asked just to return a digest also keep track of the size of the
data (in the IDL representation) so that they would stop at the same point
nodes doing full data queries would. Moreover, data queries are not
affected by per-shard memory limit and the coordinator sends individual
result size limits to replicas in order not to depend on hardcoded values.
It is still possible to get digest mismatches if the IDL changes (e.g. a
new field is added), but, hopefully, that won't be a serious problem."
* 'pdziepak/short-read-fixes/v4' of github.com:cloudius-systems/seastar-dev:
query: introduce result_memory_accounter::foreign_state
storage_proxy: fix short reads in parallel range queries
storage_proxy: pass maximum result size to replicas
mutation_partition: use result limiter for digest reads
query: make result_memory_limiter constants available for linker
result_memory_limiter: add accounter for digest reads
idl: allow writers to use any output stream
result_memory_limiter: split new_read() to new_{data, mutation}_read()
idl: is_short_read() was added in 1.6
mutation_partition: honour allowed_short_read for static rows
storage_proxy: fix _is_short_read computation
storage_proxy: disallow short reads if got no live rows
storage_proxy: don't stop after result with no live rows
This reverts commit aa392810ff, reversing
changes made to a24ff47c637e6a5fd158099b8a65f1191fc2d023; it uses
boost::intrusive::detail directly, which it must not, and doesn't compile on
all boost versions as a consequence.
* seastar 0b98024...f32e4c2 (11):
> Merge "Moving the reactor counters to the metric layer" from Amnon
> metrics: Metrics function should take variable as a refernce
> Revert "Merge ""Moving the reactor counters to the metric layer from Amnon"
> Merge ""Moving the reactor counters to the metric layer from Amnon
> Revert "fstream: Auto-close data_sink and data_source"
> rpc: Avoid resource unit leaks on failure
> fstream: Auto-close data_sink and data_source
> http: Move metrics registration to the metrics layer
> output_stream: add batching to zero copy interface
> Revert "slab: Move the metrics registration to the metrics layer"
> slab: Move the metrics registration to the metrics layer
"Reduce the size of mutation_partition by implementing intrusive set using
bi::rbtree_algorithms directly and using tree nodes optimized for size.
This will reduce the size of mutation_partition by:
24 bytes + <number of cql rows> * 8 bytes
This should have a positive impact on performance because mutation_partitions
are stored both in memtable and cache.
Fixes #742."
* 'haaawk/742' of github.com:cloudius-systems/seastar-dev:
intrusive_set: rename size() to calculate_size()
Make intrusive_set_external_comparator::_value_traits static
Implement intrusive set using rbtree_algorithms
mutation_partition: make apply_reversibly_intrusive_set nongeneric
mutation_partition: take schema in find_row and clustered_row
mutation_partition: Extract intrusive set logic to a class.
mutation_partition: Replace value_comp with key_comp calls
This hopefully will make it more apparent that
the time complexity of this method is O(N) not O(1).
Signed-off-by: Piotr Jastrzebski <piotr@scylladb.com>
_value_traits can be shared among all instances
and there's no need to store it in every single one.
Signed-off-by: Piotr Jastrzebski <piotr@scylladb.com>
This new implementation takes less memory because it
does not store comparator.
It also uses tree nodes optimized for size. This means
that instead of storing an enum field |color| they embed
this information inside pointer to parent.
Signed-off-by: Piotr Jastrzebski <piotr@scylladb.com>
apply_reversibly_intrusive_set is used only in one place
and always with rows_type. There's no need for it to be generic.
This will allow changing intrusive set implementation.
Signed-off-by: Piotr Jastrzebski <piotr@scylladb.com>
file output streams take the responsibility of closing the file, they
will close the file as part of closing the stream.
During sstable writing we create sstable object and keep file
references there as well. Sstable object also has responsibility for
closing the files, and does so from sstable::~sstable().
Double close was supposed to be avoided by a construct like this:
writer.close().get();
_file = {};
However if close() failed, which can happen when write-ahead failed,
_file would not be cleared, and both the writer and sstable would
close the file. This will result in a crash in
append_challenged_posix_file_impl::close(), which is not prepared to
be closed twice.
Another problem is that if exception happened before we reached that
construct, we still should close the writer. Currently we don't, so
there's no double close on the file, but that's a bug which needs to
be fixed and once that's fixed double close on _file will be even more
likely.
The fix employed here is to not keep files inside sstable object when
writing. As soon as the writer is constructed, it's the only owner of
the file.
Fixes#1764.
Message-Id: <1482428648-22553-1-git-send-email-tgrabiec@scylladb.com>
Range queries used to be performed sequentially and the shard performing
part of the read was reading state of the merger's memory accounter
directly. Now, they may be performed in parallel so it is safer to just
pass relevant data by value to the intersted shards so that they are not
reading something that another shard is modyfing at the same time.
Since query is done in parallel there is a chance of overread. However,
the parallelism is high only in sparsely populated tables and that's
when the overread is less serious problem.
Since a1cafed370 "storage_proxy: handle
range scans of sparsely populated tables" nonsingular range queries may
be performed in parallel on multiple shards. The consequence of this
that result may be added to the merger out of order. This requires more
complex logic for handling short reads.
As soon as mutation_result_merger gets a short read it starts to discard
all subsequently received results that are known to contain partitions
with larger keys.
Then when the final result is being prepared the merger may need to
combine and sorts results which ordering is not known. If at least one
of these results is a short one all partitions with larger keys are
removed.
Due to request being performed in parallel it is possible that even
though there was a short read the merger has got enough live data to
satisfy specified limits. If this has happened the short read flag is
not set on the final result.
We may want to change the default individual result size limit in the
future. If it is provided by the coordinator and not hardcoded in the
replicas this can be done without causing data query digest mismatches
or wasteful mutation query results.
Even if we are performing a digest query we should do proper result
memory accounting so that the result ends exactly in the same place that
it would if it was a data query. This is to avoid digest mismatches
between replicas.
It was the observation that ring_position_range_sharder doesn't support
wrapping ranges that started the nonwrapping_range madness, but that
class still has some leftover wrapping ranges. Close the circle by
removing them.
Message-Id: <20161123153113.8944-1-avi@scylladb.com>
Digest reads differ from data reads in a way that they do not really
consume any memory. We still want them to stop in the same place that
data reads would, but the per-shard semaphore shouldn't be updated by
them.
Original IDL generated code was hardcoded to always use bytes_ostream.
This patch makes the output stream a template parameter so that any
valid output stream can be used.
Unfortunately, making IDL writers generic requires updates in the code
that uses them, this is fixed in C++17 which would be able to deduce the
parameter in most cases.
For data queries it is very important that all replicas get limited in
the same place (this includes replicas returning only digest). That's
why they shouldn't be affected by per-shard result memory limit.
Moreover, we should make sure that individual memory limits are the
same, making the coordinator provide it for replicas which allow to
safely change it in the future.
Mutation queries are not as sensitive but it is still beneficial to make
sure that all replicas use the same individual limit.
If after reconciliation the coordinator ends up with no live rows and
short reads are allowed a retry may not make any progress if replicas
end their reads in the same place. The solution is to disallow short
reads on retries which are caused by final result having no live rows.
mutation_result_merger merges results from different shards and stops as
soon as a shard returned a short read or memory usage on the merging
shard is too high. However, it should never stop unless at least one
live rows is in the merged result.
* 'tgrabiec/calculate-hash-once-compaction' of github.com:cloudius-systems/seastar-dev:
sstables: Calculate key hash only once during compaction
tests: sstables: Add more test cases to tombstone_purge_test
db: Expose column_family::add_sstable
tests: sstables: Ensure timestamps are increasing
tests: sstables: Simplify tombstone_purge_test
- moved to seastar thread
- extracted sstable creation and validation logic
- reduced code duplication
- switched to mutation_reader assertions
- used result of compact_sstable() to locate the new sstable
- rather than setting gc timestamp in the past, bump the clock
before compacting
Some code paths were obtaining db_clock timestamp to only convert it
to gc_clock later. Avoid this. In the future we could make gc_clock
cheaper cause it has low precision.
Message-Id: <1482401190-2035-1-git-send-email-tgrabiec@scylladb.com>
Shared sstables will now be resharded in the same order to guarantee
that all shards owning a sstable will agree on its deletion nearly
the same time, therefore, reducing disk space requirement.
That's done by picking which column family to reshard in UUID order,
and each individual column family will reshard its shared sstables
in generation order.
Fixes#1952.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Message-Id: <87ff649ed24590c55c00cbb32bffd8fa2743e36e.1482342754.git.raphaelsc@scylladb.com>
Currently server-side timestamps use a clock with millisecond
precision. Timestamps have microsecond resolution, with lower bits
used to serialize mutations originating from given client.
Timestamps for column drops always use just the millisecond base. A
column drop which is executed after an insert may thus be given lower
timestamp than the insert, even when the two are serialized on the
client side over same connection.
Use microsecond precision to reduce chances of that event.
This is supposed to fix sporadic failures of
schema_test.py:TestSchema.drop_column_queries_test dtest.
Message-Id: <1482343119-27698-1-git-send-email-tgrabiec@scylladb.com>
Refs #1943.
* 'tgrabiec/optimize-bloom-filter' of github.com:cloudius-systems/seastar-dev:
db: Compute key hash once in partition_presence_checker
bloom_filter: Allow checking presence using pre-hashed key
db: Use incremental selector in partition_presence_checker
RPC messaging service is initialized before the Tracing service, so
we should prevent creation of tracing spans before the service is
fully initialized.
We will use an already existing "_down" state and extend it in a way
that !_down equals "started", where "started" is TRUE when the local
service is fully initialized.
We will also split the Tracing service initialization into two parts:
1) Initialize the sharded object.
2) Start the tracing service:
- Create the I/O backend service.
- Enable tracing.
Fixes issue #1939
Signed-off-by: Vlad Zolotarov <vladz@scylladb.com>
Message-Id: <1481836429-28478-1-git-send-email-vladz@scylladb.com>
If a node gets more MUTATION request that it can handle via RPC it will
stop reading from this RPC connection, but this will prevent it from
getting MUTATION_DONE responses for requests it coordinates because
currently MUTATION and MUTATION_DONE messages shares same connection.
To solve this problem this patches moves MUTATION_DONE messages to
separate connection.
Fixes: #1843
Message-Id: <20161201155942.GC11581@scylladb.com>
