LSA memory reclaimer logic assumes that the amount of memory used by LSA
equals: segments_in_use * segment_size. However, LSA is also responsible
for eviction of large objects which do not affect the used segmentcount,
e.g. region with no used segments may still use a lot of memory for
large objects. The solution is to switch from measuring memory in used
segments to used bytes count that includes also large objects.
Signed-off-by: Paweł Dziepak <pdziepak@scylladb.com>
Since this won't check disk types, may re-initialize RAID on EBS when first block was lost.
But in such condition, probably re-initialize RAID is the only choice we can take, so this is fine.
Fixes#364.
Signed-off-by: Takuya ASADA <syuu@scylladb.com>
With this patch, start two nodes
node 1:
scylla --rpc-address 127.0.0.1 --broadcast-rpc-address 127.0.0.11
node 2:
scylla --rpc-address 127.0.0.2 --broadcast-rpc-address 127.0.0.12
On node 1:
cqlsh> SELECT rpc_address from system.peers;
rpc_address
-------------
127.0.0.12
which means client should use this address to connect node 2 for cql and
thrift protocol.
It is same as
-Dcassandra.consistent.rangemovement
in cassandra.
Use it as:
$ scylla --consistent-rangemovement 0
or
$ scylla --consistent-rangemovement 1
Messaging service closes connection in rpc call continuation on
closed_error, but the code runs for each outstanding rpc call on the
connection, so first continuation may destroy genuinely closed connection,
then connection is reopened and next continuation that handless previous
error kills now perfectly healthy connection. Fix this by closing
connection only in error state.
From Avi:
Origin supports a notion of empty values for non-container types; these
are serialized as zero-length blobs. They are mostly useless and only
retained for compatibility.
The implementation here introduces a wrapper maybe_empty<T>, similar to
optional<T> but oriented towards usually-nonempty usage with implicit
conversion.
There is more work needed for full empty support: fixing up deserializers to
create empty values instead of nulls, and splitting up data_value into
data_value and a data_value_nonnull for the cases that require it.
(I chose maybe_empty<> rather than using optional<data_value> for nullable
data_value both because it requires fewer changes, and because
optional<data_value> introduces a lot of control flow when moving or copying,
which would be mostly useless in most cases).
This cleanup patch got lost in git-space some time ago. It is however sorely
needed...
* Use cleaner wrapper for creating temp dir + commit log, avoiding
having to clear and clean in every test, etc.
* Remove assertions based on file system checks, since these are not
valid due to both the async nature of the CL, and more to the point,
because of pre-allocation of files and file blocks. Use CL
counters/methods instead
* Fix some race conditions to ensure tests are safe(r)
* Speed up some tests
Discern fatal and non-fatal excceptions, and handle data corruption
by adding to stats, resporting it, but continue processing.
Note that "invalid_arguement", i.e. attempting to replay origin/old
segments are still considered fatal, as it is probably better to
signal this strongly to user/admin
Parser object now attempts to skip past/terminate parsing on corrupted
entries/chunks (as detected by invalid sizes/crc:s). The amount of data
skipped is kept track of (as well as we can estimate - pre-allocation
makes it tricky), and at the end of parsing/reporting, IFF errors
occurred, and exception detailing the failures is thrown (since
subsciption has little mechanism to deal with this otherwise).
Thus a caller can decide how to deal with data corruption, but will be
given as many entries as possible.
An empty serialized representation means an empty value, not NULL.
Fix up the confusion by converting incorrect make_null() calls to a new
make_empty(), and removing make_null() in empty-capable types like
bytes_type.
Collections don't support empty serialized representations, so remove
the call there.
Paramter evaluation order is unspecified, so it's possible that the
move of 'schema' into lambda captures would happen before construction of
mutation.
"To speed up boot, parallelism was introduced to our code that loads
sstables from a column family, a function was implemented to read
the minimum from a sstable to determine whether it belongs to the
current shard, and buffer size in read simple is dynamically chosen
based on the size of the file and dma alignment.
The latter is important because filter file can be considerably
large when the respective sstable (data file) is very large.
Before this patchset, scylla took about 5 minutes to boot with a
data directory of 660GB. After this patchset, scylla took about 20
seconds to boot with the same data directory."
Avi says:
"A small buffer size will hurt if we read a large file, but
a large buffer size won't hurt if we read a small file, since
we close it immediately."
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Currently, we only determine if a sstable belongs to current shard
after loading some of its components into memory. For example,
filter may be considerably big and its content is irrelevant to
decide if a sstable should be included to a given shard.
Start using the functions previously introduced to optimize the
sstable loading process. add_sstable no longer checks if a sstable
is relevant to the current shard.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Boot may be slow because the function that loads sstables do so
serially instead of in parallel. In the callback supplied to
lister::scan_dir, let's push the future returned by probe_file
(function that loads sstable) into a vector of future and wait
for all of them at the end.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
We cannot share some dependency package names between 14.04 and 15.10, so need to add ifdefs.
Not tested on other version of Ubuntu.
Signed-off-by: Takuya ASADA <syuu@scylladb.com>
Origin supports (https://issues.apache.org/jira/browse/CASSANDRA-5648) "empty"
values even for non-container types such as int. Use maybe_empty<> to
encapsulate abstract_type::native_type, adding an empty flag if needed.
Similar to optional<>, with the following differences:
- decays back to the encapsulated type, with an emptiness check;
this reflects the expectation that the value will rarely be empty
- avoids conditionals during copy/move (and requires a default constructor),
again with the same expectation.
When we start to sending mutations for cf_id to remote node, remote node
might do not have the cf_id anymore due to dropping of the cf for
instance.
We should not fail the streaming if this happens, since the cf does not
exist anymore there is no point streaming it.
Fixes#566
When a new node joins a cluster, it will starts a gossip round with seed
node. However, within this round, the seed node will not tell the new
node anything it knows about other nodes in the cluster, because the
digest in the gossip SYN message contains only the new node itself and
no other nodes. The seed node will pick randomly from the live nodes,
including the newly added node in do_gossip_to_live_member to start a
gossip round. If the new node is "lucky", seed node will talk to it very
soon and tells all the information it knows about the cluster, thus the
new node will mark the seed node alive and think it has seen the seed
node. If there considerably large number of live nodes, it might take a
long time before the seed node pick the new node and talk to it.
In bootstrap code, storage_service::bootstrap checks if we see any nodes
after sleep of RING_DELAY milliseconds and throw "Unable to contact any
seeds!" if not, thus the node will fail to bootstrap.
To help the seed node talk to new node faster, we favor new node in
do_gossip_to_live_member.
In origin, get_all_endpoint_states perform all the information
formatting and returns a string.
This is not a good API approach, this patch replaces the implementation
so the API will return an array of values and the JMX will do the
formatting.
This is a better API and would make it simpler in the future to stay in
sync with origin output.
This patch is part of #508
Signed-off-by: Amnon Heiman <amnon@scylladb.com>
Fixes#551.
Change mountpoint to /var/lib/scylla, copy conf/ on it.
Note: need to replace conf/ with symlink to /etc/scylla when new rpm uploaded on yum repository.
Signed-off-by: Takuya ASADA <syuu@scylladb.com>
Signed-off-by: Pekka Enberg <penberg@iki.fi>
