In another patch I noticed gcc producing dead functions. I am not sure
why gcc is doing that. Some of those functions are already placed in
independent sections, and so can be garbage collected by the linker.
This is a 1% text section reduction in scylla, from 39363380 to
38974324 bytes. There is no difference in the tps reported by
perf_simple_query.
Signed-off-by: Rafael Ávila de Espíndola <espindola@scylladb.com>
Message-Id: <20200723152511.8214-1-espindola@scylladb.com>
In the patch "Add exception overloads for Dynamo types", Alternator's single
api_error exception type was replaced by a more complex hierarchy of types.
The implementation was not only longer and more complex to understand -
I believe it also negated an important observation:
The "api_error" exception type is special. It is not an exception created
by code for other code. It is not meant to be caught in Alternator code.
Instead, it is supposed to contain an error message created for the *user*,
containing one of the few supported exception exception "names" described
in the DynamoDB documentation, and a user-readable text message. Throwing
such an exception in Alternator code means the thrower wants the request
to abort immediately, and this message to reach the user. These exceptions
are not designed to be caught in Alternator code. Code should use other
exceptions - or alternatives to exceptions (e.g., std::optional) for
problems that should be handled before returning a different error to the
user. Moreover, "api_error" isn't just thrown as an exception - it can
also be returned-by-value in a executor::request_return_type) - which is
another reason why it should not be subclassed.
For these reasons, I believe we should have a single api_error type, and
it's wrong to subclass it. So in this patch I am reverting the subclasses
and template added in the aforementioned patch.
Still, one correct observation made in that patch was that it is
inconvenient to type in DynamoDB exception names (no help from the editor
in completing those strings) and also error-prone. In this patch we
propse a different - simpler - solution to the same problem:
We add trivial factory functions, e.g., api_error::validation(std::string)
as a shortcut to api_error("ValidationException"). The new implementation
is easy to understand, and also more self explanatory to readers:
It is now clear that "api_error::validation()" is actually a user-visible
"api_error", something which was obscured by the name validation_exception()
used before this patch.
Finally, this patch also improves the comment in error.hh explaining the
purpose of api_error and the fact it can be returned or thrown. The fact
it should not be subclassed is legislated with a "finally". There is also
no point of this class inheriting from std::exception or having virtual
functions, or an empty constructor - so all these are dropped as well.
Signed-off-by: Nadav Har'El <nyh@scylladb.com>
Ninja has a special pool called console that causes programs in that
pool to output directly to the console instead of being logged. By
putting test.py in it is now possible to run just
$ ninja dev-test
And see the test.py output while it is running.
Signed-off-by: Rafael Ávila de Espíndola <espindola@scylladb.com>
Message-Id: <20200716204048.452082-1-espindola@scylladb.com>
The targets {dev|debug|release}-test run all unit tests, including
alternator/run. But this test requires the Scylla executable, which
wasn't among the dependencies. Fix it by adding build/$mode/scylla to
the dependency list.
Fixes#6855.
Tests: `ninja dev-test` after removing build/dev/scylla
Signed-off-by: Dejan Mircevski <dejan@scylladb.com>
Merged pull request https://github.com/scylladb/scylla/pull/6694
by Calle Wilund:
Implementation of DynamoDB streams using Scylla CDC.
Fixes#5065
Initial, naive implementation insofar that it uses 1:1 mapping CDC stream to
DynamoDB shard. I.e. there are a lot of shards.
Includes tests verified against both local DynamoDB server and actual AWS
remote one.
Note:
Because of how data put is implemented in alternator, currently we do not
get "proper" INSERT labels for first write of data, because to CDC it looks
like an update. The test compensates for this, but actual users might not
like it.
The data model is now
bplus::tree<Key = int64_t, T = array<entry>>
where entry can be cache_entry or memtable_entry.
The whole thing is encapsulated into a collection called "double_decker"
from patch #3. The array<T> is an array of T-s with 0-bytes overhead used
to resolve hash conflicts (patch #2).
branch:
tests: unit(debug)
tests before v7:
unit(debug) for new collections, memtable and row_cache
unit(dev) for the rest
perf(dev)
* https://github.com/xemul/scylla/commits/row-cache-over-bptree-9:
test: Print more sizes in memory_footprint_test
memtable: Switch onto B+ rails
row_cache: Switch partition tree onto B+ rails
memtable: Count partitions separately
token: Introduce raw() helper and raw comparator
row-cache: Use ring_position_comparator in some places
dht: Detach ring_position_comparator_for_sstables
double-decker: A combination of B+tree with array
intrusive-array: Array with trusted bounds
utils: B+ tree implementation
test: Move perf measurement helpers into header
Add types exception overloads for ValidationException, ResourceNotFoundException, etc,
to avoid writing explicit error type as string everywhere (with the potential for
spelling errors ever present).
Also allows intellisense etc to complete the exception when coded.
The collection is K:V store
bplus::tree<Key = K, Value = array_trusted_bounds<V>>
It will be used as partitions cache. The outer tree is used to
quickly map token to cache_entry, the inner array -- to resolve
(expected to be rare) hash collisions.
It also must be equipped with two comparators -- less one for
keys and full one for values. The latter is not kept on-board,
but it required on all calls.
The core API consists of just 2 calls
- Heterogenuous lower_bound(search_key) -> iterator : finds the
element that's greater or equal to the provided search key
Other than the iterator the call returns a "hint" object
that helps the next call.
- emplace_before(iterator, key, hint, ...) : the call construct
the element right before the given iterator. The key and hint
are needed for more optimal algo, but strictly speaking not
required.
Adding an entry to the double_decker may result in growing the
node's array. Here to B+ iterator's .reconstruct() method
comes into play. The new array is created, old elements are
moved onto it, then the fresh node replaces the old one.
// TODO: Ideally this should be turned into the
// template <typename OuterCollection, typename InnerCollection>
// but for now the double_decker still has some intimate knowledge
// about what outer and inner collections are.
Insertion into this collection _may_ invalidate iterators, but
may leave intact. Invalidation only happens in case of hashing
conflict, which can be clearly seen from the hint object, so
there's a good room for improvement.
The main usage by row_cache (the find_or_create_entry) looks like
cache_entry find_or_create_entry() {
bound_hint hint;
it = lower_bound(decorated_key, &hint);
if (!hint.found) {
it = emplace_before(it, decorated_key.token(), hint,
<constructor args>)
}
return *it;
}
Now the hint. It contains 3 booleans, that are
- match: set to true when the "greater or equal" condition
evaluated to "equal". This frees the caller from the need
to manually check whether the entry returned matches the
search key or the new one should be inserted.
This is the "!found" check from the above snippet.
To explain the next 2 bools, here's a small example. Consider
the tree containing two elements {token, partition key}:
{ 3, "a" }, { 5, "z" }
As the collection is sorted they go in the order shown. Next,
this is what the lower_bound would return for some cases:
{ 3, "z" } -> { 5, "z" }
{ 4, "a" } -> { 5, "z" }
{ 5, "a" } -> { 5, "z" }
Apparently, the lower bound for those 3 elements are the same,
but the code-flows of emplacing them before one differ drastically.
{ 3, "z" } : need to get previous element from the tree and
push the element to it's vector's back
{ 4, "a" } : need to create new element in the tree and populate
its empty vector with the single element
{ 5, "a" } : need to put the new element in the found tree
element right before the found vector position
To make one of the above decisions the .emplace_before would need
to perform another set of comparisons of keys and elements.
Fortunately, the needed information was already known inside the
lower_bound call and can be reported via the hint.
Said that,
- key_match: set to true if tree.lower_bound() found the element
for the Key (which is token). For above examples this will be
true for cases 3z and 5a.
- key_tail: set to true if the tree element was found, but when
comparing values from array the bounding element turned out
to belong to the next tree element and the iterator was ++-ed.
For above examples this would be true for case 3z only.
And the last, but not least -- the "erase self" feature. Which is
given only the cache_entry pointer at hands remove it from the
collection. To make this happen we need to make two steps:
1. get the array the entry sits in
2. get the b+ tree node the vectors sits in
Both methods are provided by array_trusted_bounds and bplus::tree.
So, when we need to get iterator from the given T pointer, the algo
looks like
- Walk back the T array untill hitting the head element
- Call array_trusted_bounds::from_element() getting the array
- Construct b+ iterator from obtained array
- Construct the double_decker iterator from b+ iterator and from
the number of "steps back" from above
- Call double_decker::iterator.erase()
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
A plain array of elements that grows and shrinks by
constructing the new instance from an existing one and
moving the elements from it.
Behaves similarly to vector's external array, but has
0-bytes overhead. The array bounds (0-th and N-th
elemements) are determined by checking the flags on the
elements themselves. For this the type must support
getters and setters for the flags.
To remove an element from array there's also a nothrow
option that drops the requested element from array,
shifts the righter ones left and keeps the trailing
unused memory (so called "train") until reconstruction
or destruction.
Also comes with lower_bound() helper that helps keeping
the elements sotred and the from_element() one that
returns back reference to the array in which the element
sits.
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
// The story is at
// https://groups.google.com/forum/#!msg/scylladb-dev/sxqTHM9rSDQ/WqwF1AQDAQAJ
This is the B+ version which satisfies several specific requirements
to be suitable for row-cache usage.
1. Insert/Remove doesn't invalidate iterators
2. Elements should be LSA-compactable
3. Low overhead of data nodes (1 pointer)
4. External less-only comparator
5. As little actions on insert/delete as possible
6. Iterator walks the sorted keys
The design, briefly is:
There are 3 types of nodes: inner, leaf and data, inner and leaf
keep build-in array of N keys and N(+1) nodes. Leaf nodes sit in
a doubly linked list. Data nodes live separately from the leaf ones
and keep pointers on them. Tree handler keeps pointers on root and
left-most and right-most leaves. Nodes do _not_ keep pointers or
references on the tree (except 3 of them, see below).
changes in v9:
- explicitly marked keys/kids indices with type aliases
- marked the whole erase/clear stuff noexcept
- disposers now accept object pointer instead of reference
- clear tree in destructor
- added more comments
- style/readability review comments fixed
Prior changes
**
- Add noexcepts where possible
- Restrict Less-comparator constraint -- it must be noexcept
- Generalized node_id
- Packed code for beging()/cbegin()
**
- Unsigned indices everywhere
- Cosmetics changes
**
- Const iterators
- C++20 concepts
**
- The index_for() implmenetation is templatized the other way
to make it possible for AVX key search specialization (further
patching)
**
- Insertion tries to push kids to siblings before split
Before this change insertion into full node resulted into this
node being split into two equal parts. This behaviour for random
keys stress gives a tree with ~2/3 of nodes half-filled.
With this change before splitting the full node try to push one
element to each of the siblings (if they exist and not full).
This slows the insertion a bit (but it's still way faster than
the std::set), but gives 15% less total number of nodes.
- Iterator method to reconstruct the data at the given position
The helper creates a new data node, emplaces data into it and
replaces the iterator's one with it. Needed to keep arrays of
data in tree.
- Milli-optimize erase()
- Return back an iterator that will likely be not re-validated
- Do not try to update ancestors separation key for leftmost kid
This caused the clear()-like workload work poorly as compared to
std:set. In particular the row_cache::invalidate() method does
exactly this and this change improves its timing.
- Perf test to measure drain speed
- Helper call to collect tree counters
**
- Fix corner case of iterator.emplace_before()
- Clean heterogenous lookup API
- Handle exceptions from nodes allocations
- Explicitly mark places where the key is copied (for future)
- Extend the tree.lower_bound() API to report back whether
the bound hit the key or not
- Addressed style/cleanness review comments
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
The default ninja build target now builds artifacts and packages. Let's
add a 'build' target that only builds the artifacts.
Message-Id: <20200714105042.416698-1-penberg@scylladb.com>
This adds a new "dist-<mode>" target, which builds the server package in
selected build mode together with the other packages, and wires it to
the "<mode>" target, which is built as part of default "ninja"
invocation.
This allows us to perform a full build, package, and test cycle across
all build modes with:
./configure.py && ninja && ./test.py
Message-Id: <20200713101918.117692-1-penberg@scylladb.com>
"
This is the first stage of replacing the existing restrictions code with a new representation. It adds a new class `expression` to replace the existing class `restriction`. Lots of the old code is deleted, though not all -- that will come in subsequent stages.
Tests: unit (dev, debug restrictions_test), dtest (next-gating)
"
* dekimir-restrictions-rewrite:
cql3/restrictions: Drop dead code
cql3/restrictions: Use free functions instead of methods
cql3/restrictions: Create expression objects
cql3/restrictions: Add free functions over new classes
cql3/restrictions: Add new representation
Instead of `restriction` class methods, use the new free functions.
Specific replacement actions are listed below.
Note that class `restrictions` (plural) remains intact -- both its
methods and its type hierarchy remain intact for now.
Ensure full test coverage of the replacement code with new file
test/boost/restrictions_test.cc and some extra testcases in
test/cql/*.
Drop some existing tests because they codify buggy behaviour
(reference #6369, #6382). Drop others because they forbid relation
combinations that are now allowed (eg, mixing equality and
inequality, comparing to NULL, etc.).
Here are some specific categories of what was replaced:
- restriction::is_foo predicates are replaced by using the free
function find_if; sometimes it is used transitively (see, eg,
has_slice)
- restriction::is_multi_column is replaced by dynamic casts (recall
that the `restrictions` class hierarchy still exists)
- utility methods is_satisfied_by, is_supported_by, to_string, and
uses_function are replaced by eponymous free functions; note that
restrictions::uses_function still exists
- restriction::apply_to is replaced by free function
replace_column_def
- when checking infinite_bound_range_deletions, the has_bound is
replaced by local free function bounded_ck
- restriction::bounds and restriction::value are replaced by the more
general free function possible_lhs_values
- using free functions allows us to simplify the
multi_column_restriction and token_restriction hierarchies; their
methods merge_with and uses_function became identical in all
subclasses, so they were moved to the base class
- single_column_primary_key_restrictions<clustering_key>::needs_filtering
was changed to reuse num_prefix_columns_that_need_not_be_filtered,
which uses free functions
Fixes#5799.
Fixes#6369.
Fixes#6371.
Fixes#6372.
Fixes#6382.
Signed-off-by: Dejan Mircevski <dejan@scylladb.com>
Merged patch series by Piotr Sarna:
The alternator project was in need of a more optimized
JSON library, which resulted in creating "rjson" helper functions.
Scylla generally used libjsoncpp for its JSON handling, but in order
to reduce the dependency hell, the usage is now migrated
to rjson, which is faster and offers the same functionality.
The original plan was to be able to drop the dependency
on libjsoncpp-lib altogether and remove it from install-dependencies.sh,
but one last usage of it remains in our test suite,
namely cql_repl. The tool compares its output JSON textually,
so it depends on how a library presents JSON - what are the delimeters,
indentation, etc. It's possible to provide a layer of translation
to force rjson to print in an identical format, but the other issue
is that libjsoncpp keeps subobjects sorted by their name,
while rjson uses an unordered structure.
There are two possible solutions for the last remaining usage
of libjsoncpp:
1. change our test suite to compare JSON documents with a JSON parser,
so that we don't rely on internal library details
2. provide a layer of translation which forces rjson to print
its objects in a format idential to libjsoncpp.
(1.) would be preferred, since now we're also vulnerable for changes
inside libjsoncpp itself - if they change anything in their output
format, tests would start failing. The issue is not critical however,
so it's left for later.
Tests: unit(dev), manual(json_test),
dtest(partitioner_tests.TestPartitioner.murmur3_partitioner_test)
Piotr Sarna (8):
alternator,utils: move rjson.hh to utils/
alternator: remove ambiguous string overloads in rjson
rjson: add parse_to_map helper function
rjson: add from_string_map function
rjson: add non-throwing parsing
rjson: move quote_json_string to rjson
treewide: replace libjsoncpp usage with rjson
configure: drop json.cc and json.hh helpers
alternator/base64.hh | 2 +-
alternator/conditions.cc | 2 +-
alternator/executor.hh | 2 +-
alternator/expressions.hh | 2 +-
alternator/expressions_types.hh | 2 +-
alternator/rmw_operation.hh | 2 +-
alternator/serialization.cc | 2 +-
alternator/serialization.hh | 2 +-
alternator/server.cc | 2 +-
caching_options.hh | 9 +-
cdc/log.cc | 4 +-
column_computation.hh | 5 +-
configure.py | 3 +-
cql3/functions/functions.cc | 4 +-
cql3/statements/update_statement.cc | 24 ++--
cql3/type_json.cc | 212 ++++++++++++++++++----------
cql3/type_json.hh | 7 +-
db/legacy_schema_migrator.cc | 12 +-
db/schema_tables.cc | 1 -
flat_mutation_reader.cc | 1 +
index/secondary_index.cc | 80 +++++------
json.cc | 80 -----------
json.hh | 113 ---------------
schema.cc | 25 ++--
test/boost/cql_query_test.cc | 9 +-
test/manual/json_test.cc | 4 +-
test/tools/cql_repl.cc | 1 +
{alternator => utils}/rjson.cc | 75 +++++++++-
{alternator => utils}/rjson.hh | 40 +++++-
29 files changed, 344 insertions(+), 383 deletions(-)
delete mode 100644 json.cc
delete mode 100644 json.hh
rename {alternator => utils}/rjson.cc (86%)
rename {alternator => utils}/rjson.hh (81%)
Thrift 0.12 includes a change [1] that avoids writing the generated output
if it has not changed. As a result, if you touch cassandra.thrift
(but not change it), the generated files will not update, and as
a result ninja will try to rebuild them every time. The compilation
of thrift files will be fast due to ccache, but still we will re-link
everything.
This touching of cassandra.thrift can happen naturally when switching
to a different git branch and then switching back. The net result
is that cassandra.thrift's contents has not changed, but its timestamp
has.
Fix by adding the "restat" option to the thrift rule. This instructs
ninja to check of the output has changed as expected or not, and to
avoid unneeded rebuilds if it has not.
[1] https://issues.apache.org/jira/browse/THRIFT-4532
This patch set adds a few new features in order to fix issue
The list of changes is briefly as follows:
- Add a new `LWT` flag to `cql3::prepared_metadata`,
which allows clients to clearly distinguish betwen lwt and
non-lwt statements without need to execute some custom parsing
logic (e.g. parsing the prepared query with regular expressions),
which is obviously quite fragile.
- Introduce the negotiation procedure for cql protocol extensions.
This is done via `cql_protocol_extension` enum and is expected
to have an appropriate mirroring implementation on the client
driver side in order to work properly.
- Implmenent a `LWT_ADD_METADATA_MARK` cql feature on top of the
aforementioned algorithm to make the feature negotiable and use
it conditionally (iff both server and client agrees with each
other on the set of cql extensions).
The feature is meant to be further utilized by client drivers
to use primary replicas consistently when dealing with conditional
statements.
* git@github.com:ManManson/scylla feature/lwt_prepared_meta_flag_2:
lwt: introduce "LWT" flag in prepared statement metadata
transport: introduce `cql_protocol_extension` enum and cql protocol extensions negotiation
"
The snapshotting code is already well isolated from the rest of
the storage_service, so it's relatively easy to move it into
independent component, thus de-bloating the storage_service.
As a side effect this allows painless removal of calls to global
get_storage_service() from schema::describe code.
Test: unit(debug), dtest.snapshot_test(dev), manual start-stop
"
* 'br-snapshot-controller-4' of https://github.com/xemul/scylla:
snap: Get rid of storage_service reference in schema.cc
main: Stop http server
snapshot: Make check_snapshot_not_exist a method
snapshots: Move ops gate from storage_service
snapshot: Move lock from storage_service
snapshot: Move all code into db::snapshot_ctl class
storage_service: Move all snapshot code into snapshot-ctl.cc
snapshots: Initial skeleton
snapshots: Properly shutdown API endpoints
api: Rewrap set_server_snapshot lambda
As HACKING.md suggests, we now require gcc version >= 10. Set the
minimum at 10.1.1, as that is the first official 10 release:
https://gcc.gnu.org/releases.html
Tests: manually run configure.py and ensure it passes/fails
appropriately.
Signed-off-by: Dejan Mircevski <dejan@scylladb.com>
A placeholder for snapshotting code that will be moved into it
from the storage_service.
Also -- pass it through the API for future use.
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
"
This patch series attempts to decouple package build and release
infrastructure, which is internal to Scylla (the company). The goal of
this series is to make it easy for humans and machines to build the full
Scylla distribution package artifacts, and make it easy to quickly
verify them.
The improvements to build system are done in the following steps.
1. Make scylla.git a super-module, which has git submodules for
scylla-jmx and scylla-tools. A clone of scylla.git is now all that
is needed to access all source code of all the different components
that make up a Scylla distribution, which is a preparational step to
adding "dist" ninja build target. A scripts/sync-submodules.sh helper
script is included, which allows easy updating of the submodules to the
latest head of the respective git repositories.
2. Make builds reproducible by moving the remaining relocatable package
specific build options from reloc/build_reloc.sh to the build system.
After this step, you can build the exact same binaries from the git
repository by using the dbuild version from scylla.git.
3. Add a "dist" target to ninja build, which builds all .rpm and .deb
packages with one command. To build a release, run:
$ ./tools/toolchain/dbuild ./configure.py --mode release
$ ./tools/toolchain/dbuild ninja-build dist
and you will now have .rpm and .deb packages to all the components of
a Scylla distribution.
4. Add a "dist-check" target to ninja build for verification of .rpm and
.deb packages in one command. To verify all the built packages, run:
$ ninja-build dist-check
Please note that you must run this step on the host, because the
target uses Docker under the hood to verify packages by installing
them on different Linux distributions.
Currently only CentOS 7 verification is supported.
All these improvements are done so that backward compatibility is
retained. That is, any existing release infrastructure or other build
scripts are completely unaffacted.
Future improvements to consider:
- Package repository generation: add a "ninja repo" command to generate
a .rpm and .deb repositories, which can be uploaded to a web site.
This makes it possible to build a downloadable Scylla distribution
from scylla.git. The target requires some configuration, which user
has to provide. For example, download URL locations and package
signing keys.
- Amazon Machine Image (AMI) support: add a "ninja ami" command to
simplify the steps needed to generate a Scylla distribution AMI.
- Docker image support: add a "ninja docker" command to simplify the
steps needed to generate a Scylla distribution Docker image.
- Simplify and unify package build: simplify and unify the various shell
scripts needed to build packages in different git repositories. This
step will break backward compatiblity and can be done only after
relevant build scripts and release infrastructure is updated.
"
* 'penberg/packaging/v5' of github.com:penberg/scylla:
docs: Update packaging documentation
build: Add "dist-check" target
scripts/testing: Add "dist-check" for package verification
build: Add "dist" target
reloc: Add '--builddir' option to build_deb.sh
build: Add "-ffile-prefix-map" to cxxflags
docs: Document sync-submodules.sh script in maintainer.md
sync-submodules.sh: Add script for syncing submodules
Add scylla-tools submodule
Add scylla-jmx submodule
"
The "promoted index" is how the sstable format calls the clustering key index within a given partition.
Large partitions with many rows have it. It's embedded in the partition index entry.
Currently, lookups in the promoted index are done by scanning the index linearly so the lookup
is O(N). For large partitions that's inefficient. It consumes both a lot of CPU and I/O.
We could do better and use binary search in the index. This patch series switches the mc-format
index reader to do that. Other formats use the old way.
The "mc" format promoted index has an extra structure at the end of the index called "offset map".
It's a vector of offsets of consecutive promoted index entries. This allows us to access random
entries in the index without reading the whole index.
The location of the offset entry for a given promoted index entry can be derived by knowing where
the offset vector ends in the index file, so the offset map also doesn't have to be read completely
into the memory.
The most tricky part is caching. We need to cache blocks read from the index file to amortize the
cost of binary search:
- if the promoted index fits in the 32 KiB which was read from the index when looking for
the partition entry, we don't want to issue any additional I/O to search the promoted index.
- with large promoted indexes, the last few bisections will fall into the same I/O block and we
want to reuse that block.
- we don't want the cache to grow too big, we don't want to cache the whole promoted index
as the read progresses over the index. Scanning reads may skip multiple times.
This series implements a rather simple approach which meets all the
above requirements and is not worse than the current state of affairs:
- Each index cursor has its own cache of the index file area which corresponds to promoted index
This is managed by the cached_file class.
- Each index cursor has its own cache of parsed blocks. This allows the upper bound estimation to
reuse information obtained during lower bound lookup. This estimation is used to limit
read-aheads in the data file.
- Each cursor drops entries that it walked past so that memory footprint stays O(log N)
- Cached buffers are accounted to read's reader_permit.
Later, we could have a single cache shared by many readers. For that, we need to come up with eviction
policy.
Fixes#4007.
TESTING RESULTS
* Point reads, large promoted index:
Config: rows: 10000000, value size: 2000
Partition size: 20 GB
Index size: 7 MB
Notes:
- Slicing read into the middle of partition (offset=5000000, read=1) is a clear win for the binary search:
time: 1.9ms vs 22.9ms
CPU utilization: 8.9% vs 92.3%
I/O: 21 reqs / 172 KiB vs 29 reqs / 3'520 KiB
It's 12x faster, CPU utilization is 10x times smaller, disk utilization is 20x smaller.
- Slicing at the front (offset=0) is a mixed bag.
time is similar: 1.8ms
CPU utilization is 6.7x smaller for bsearch: 8.5% vs 57.7%
disk bandwidth utilization is smaller for bsearch but uses more IOs: 4 reqs / 320 KiB (scan) vs 17 reqs / 188 KiB (bsearch)
bsearch uses less bandwidth because the series reduces buffer size used for index file I/O.
scan is issuing:
2 * 128 KB (index page)
2 * 32 KB (data file)
bsearch is issuing:
1 * 64 KB (index page)
15 * 4 KB (promoted index)
1 * 64 KB (data file)
The 1 * 64 KB is chosen dynamically by seastar. Sometimes it chooses 2 * 32 KB (with read-ahead).
32 KB is the minimum I/O currently.
Disk utilization could be further improved by changing the way seastar's dynamic I/O adjustments work
so that it uses 1 * 4 KB when it suffices. This is left for the follow-up.
Command:
perf_fast_forward --datasets=large-part-ds1 \
--run-tests=large-partition-slicing-clustering-keys -c1 --test-case-duration=1
Before:
offset read time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
0 1 0.001836 172 1 545 9 563 175 4.0 4 320 2 2 0 1 1 0 0 0 57.7% 0
0 32 0.001858 502 32 17220 126 17776 11526 3.2 3 324 2 1 0 1 1 0 0 0 56.4% 0
0 256 0.002833 339 256 90374 427 91757 85931 7.0 7 776 3 1 0 1 1 0 0 0 41.1% 0
0 4096 0.017211 58 4096 237984 2011 241802 233870 66.1 66 8376 59 2 0 1 1 0 0 0 21.4% 0
5000000 1 0.022952 42 1 44 1 45 41 29.2 29 3520 22 2 0 1 1 0 0 0 92.3% 0
5000000 32 0.023052 43 32 1388 14 1414 1331 31.1 32 3588 26 2 0 1 1 0 0 0 91.7% 0
5000000 256 0.024795 41 256 10325 129 10721 9993 43.1 39 4544 29 2 0 1 1 0 0 0 86.4% 0
5000000 4096 0.038856 27 4096 105414 398 106918 103162 95.2 95 12160 78 5 0 1 1 0 0 0 61.4% 0
After (v2):
offset read time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
0 1 0.001831 248 1 546 21 581 252 17.6 17 188 2 0 0 1 1 0 0 0 8.5% 0
0 32 0.001910 535 32 16751 626 17770 13896 17.9 19 160 3 0 0 1 1 0 0 0 8.8% 0
0 256 0.003545 266 256 72207 2333 89076 62852 26.9 24 764 7 0 0 1 1 0 0 0 9.7% 0
0 4096 0.016800 56 4096 243812 524 245430 239736 83.6 83 8700 64 0 0 1 1 0 0 0 16.6% 0
5000000 1 0.001968 351 1 508 19 538 380 21.3 21 172 2 0 0 1 1 0 0 0 8.9% 0
5000000 32 0.002273 431 32 14077 436 15503 11551 22.7 22 268 3 0 0 1 1 0 0 0 8.9% 0
5000000 256 0.003889 257 256 65824 2197 81833 57813 34.0 37 652 18 0 0 1 1 0 0 0 11.2% 0
5000000 4096 0.017115 54 4096 239324 834 241310 231993 88.3 88 8844 65 0 0 1 1 0 0 0 16.8% 0
After (v1):
offset read time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
0 1 0.001886 259 1 530 4 545 261 18.0 18 376 2 2 0 1 1 0 0 0 9.1% 0
0 32 0.001954 513 32 16381 93 16844 15618 19.0 19 408 3 2 0 1 1 0 0 0 9.3% 0
0 256 0.003266 318 256 78393 1820 81567 61663 30.8 26 1272 7 2 0 1 1 0 0 0 10.4% 0
0 4096 0.017991 57 4096 227666 855 231915 225781 83.1 83 8888 55 5 0 1 1 0 0 0 15.5% 0
5000000 1 0.002353 232 1 425 2 432 232 23.0 23 396 2 2 0 1 1 0 0 0 8.7% 0
5000000 32 0.002573 384 32 12437 47 12571 429 25.0 25 460 4 2 0 1 1 0 0 0 8.5% 0
5000000 256 0.003994 259 256 64101 2904 67924 51427 37.0 35 1484 11 2 0 1 1 0 0 0 10.6% 0
5000000 4096 0.018567 56 4096 220609 448 227395 219029 89.8 89 9036 59 5 0 1 1 0 0 0 15.1% 0
* Point reads, small promoted index (two blocks):
Config: rows: 400, value size: 200
Partition size: 84 KiB
Index size: 65 B
Notes:
- No significant difference in time
- the same disk utilization
- similar CPU utilization
Command:
perf_fast_forward --datasets=large-part-ds1 \
--run-tests=large-partition-slicing-clustering-keys -c1 --test-case-duration=1
Before:
offset read time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
0 1 0.000279 470 1 3587 31 3829 478 3.0 3 68 2 1 0 1 1 0 0 0 21.1% 0
0 32 0.000276 3498 32 116038 811 122756 104033 3.0 3 68 2 1 0 1 1 0 0 0 24.0% 0
0 256 0.000412 2554 256 621044 1778 732150 559221 2.0 2 72 2 0 0 1 1 0 0 0 32.6% 0
0 4096 0.000510 1901 400 783883 4078 819058 665616 2.0 2 88 2 0 0 1 1 0 0 0 36.4% 0
200 1 0.000339 2712 1 2951 8 3001 2569 2.0 2 72 2 0 0 1 1 0 0 0 17.8% 0
200 32 0.000352 2586 32 91019 266 92427 83411 2.0 2 72 2 0 0 1 1 0 0 0 20.8% 0
200 256 0.000458 2073 200 436503 1618 453945 385501 2.0 2 88 2 0 0 1 1 0 0 0 29.4% 0
200 4096 0.000458 2097 200 436475 1676 458349 381558 2.0 2 88 2 0 0 1 1 0 0 0 29.0% 0
After (v1):
Testing slicing of large partition using clustering keys:
offset read time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
0 1 0.000278 492 1 3598 30 3831 500 3.0 3 68 2 1 0 1 1 0 0 0 19.4% 0
0 32 0.000275 3433 32 116153 753 122915 92559 3.0 3 68 2 1 0 1 1 0 0 0 22.5% 0
0 256 0.000458 2576 256 559437 2978 728075 504375 2.1 2 88 2 0 0 1 1 0 0 0 29.0% 0
0 4096 0.000506 1888 400 790064 3306 822360 623109 2.0 2 88 2 0 0 1 1 0 0 0 36.6% 0
200 1 0.000382 2493 1 2619 10 2675 2268 2.0 2 88 2 0 0 1 1 0 0 0 16.3% 0
200 32 0.000398 2393 32 80422 333 84759 22281 2.0 2 88 2 0 0 1 1 0 0 0 19.0% 0
200 256 0.000459 2096 200 435943 1608 453989 380749 2.0 2 88 2 0 0 1 1 0 0 0 30.5% 0
200 4096 0.000458 2097 200 436410 1651 455779 382485 2.0 2 88 2 0 0 1 1 0 0 0 29.2% 0
* Scan with skips, large index:
Config: rows: 10000000, value size: 2000
Partition size: 20 GB
Index size: 7 MB
Notes:
- Similar time, slightly worse for binary search: 36.1 s (scan) vs 36.4 (bsearch)
- Slightly more I/O for bsearch: 153'932 reqs / 19'703'260 KiB (scan) vs 155'651 reqs / 19'704'088 KiB (bsearch)
Binary search reads more by 828 KB and by 1719 IOs.
It does more I/O to read the the promoted index offset map.
- similar (low) memory footprint. The danger here is that by caching index blocks which we touch as we scan
we would end up caching the whole index. But this is protected against by eviction as demonstrated by the
last "mem" column.
Command:
perf_fast_forward --datasets=large-part-ds1 \
--run-tests=large-partition-skips -c1 --test-case-duration=1
Before:
read skip time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
1 1 36.103451 4 5000000 138491 38 138601 138453 153932.0 153932 19703260 153561 1 0 1 1 0 0 0 31.5% 502690
After (v2):
read skip time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
1 1 37.000145 4 5000000 135135 6 135146 135128 155651.0 155651 19704088 138968 0 0 1 1 0 0 0 34.2% 0
After (v1):
read skip time (s) iterations frags frag/s mad f/s max f/s min f/s avg aio aio (KiB) blocked dropped idx hit idx miss idx blk c hit c miss c blk cpu mem
1 1 36.965520 4 5000000 135261 30 135311 135231 155628.0 155628 19704216 139133 1 0 1 1 0 0 0 33.9% 248738
Also in:
git@github.com:tgrabiec/scylla.git sstable-use-index-offset-map-v2
Tests:
- unit (all modes)
- manual using perf_fast_forward
"
* tag 'sstable-use-index-offset-map-v2' of github.com:tgrabiec/scylla:
sstables: Add promoted index cache metrics
position_in_partition: Introduce external_memory_usage()
cached_file, sstables: Add tracing to index binary search and page cache
sstables: Dynamically adjust I/O size for index reads
sstables, tests: Allow disabling binary search in promoted index from perf tests
sstables: mc: Use binary search over the promoted index
utils: Introduce cached_file
sstables: clustered_index: Relax scope of validity of entry_info
sstables: index_entry: Introduce owning promoted_index_block_position
compound_compat: Allow constructing composite from a view
sstables: index_entry: Rename promoted_index_block_position to promoted_index_block_position_view
sstables: mc: Extract parser for promoted index block
sstables: mc: Extract parser for clustering out of the promoted index block parser
sstables: consumer: Extract primitive_consumer
sstables: Abstract the clustering index cursor behavior
sstables: index_reader: Rearrange to reduce branching and optionals
This patch adds "-ffile-prefix-map" to cxxflags for all build modes.
This has two benefits:
1, Relocatable packages no longer have any special build flags, which
makes deeper integration with the build system possible (e.g.
targets for packages).
2 Builds are now reproducible, which makes debugging easier in case you
only have a backtrace, but no artifacts. Rafael explains:
"BTW, I think I found another argument for why we should always build
with -ffile-prefix-map=.
There was user after free test failure on next promotion. I am unable
to reproduce it locally, so it would be super nice to be able to
decode the backtrace.
I was able to do it, but I had to create a
/jenkins/workspace/scylla-master/next/ directory and build from there
to get the same results as the bot."
Acked-by: Botond Dénes <bdenes@scylladb.com>
Acked-by: Nadav Har'El <nyh@scylladb.com>
Acked-by: Rafael Avila de Espindola <espindola@scylladb.com>
It is a read-through cache of a file.
Will be used to cache contents of the promoted index area from the
index file.
Currently, cached pages are evicted manually using the invalidate_*()
method family, or when the object is destroyed.
The cached_file represents a subset of the file. The reason for this
is to satisfy two requirements. One is that we have a page-aligned
caching, where pages are aligned relative to the start of the
underlying file. This matches requirements of the seastar I/O engine
on I/O requests. Another requirement is to have an effective way to
populate the cache using an unaligned buffer which starts in the
middle of the file when we know that we won't need to access bytes
located before the buffer's position. See populate_front(). If we
couldn't assume that, we wouldn't be able to insert an unaligned
buffer into the cache.
The patch introduces two new features to aid with negotiating
protocol extensions for the CQL protocol:
- `cql_protocol_extensions` enum, which holds all supported
extensions for the CQL protocol (currently contains only
`LWT_ADD_METADATA_MARK` extension, which will be mentioned
below).
- An additional mechainsm of negotiating cql protocol extensions
to be used in a client connection between a scylla server
and a client driver.
These extensions are propagated in SUPPORTED message sent from the
server side with "SCYLLA_" prefix and received back as a response
from the client driver in order to determine intersection between
the cql extensions that are both supported by the server and
acknowledged by a client driver.
This intersection of features is later determined to be a working
set of cql protocol extensions in use for the current `client_state`,
which is associated with a particular client connection.
This way we can easily settle on the used extensions set on
both sides of the connection.
Currently there is only one value: `LWT_ADD_METADATA_MARK`, which
regulates whether to set a designated bit in prepared statement
metadata indicating if the statement at hand is an lwt statement
or not (actual implementation for the feature will be in a later
patch).
Each extension can also propagate some custom parameters to the
corresponding key. CQL protocol specification allows to send
a list of values with each key in the SUPPORTED message, we use
that to pass parameters to extensions as `PARAM=VALUE` strings.
In case of `LWT_ADD_METADATA_MARK` it's
`SCYLLA_LWT_OPTIMIZATION_META_BIT_MASK` which designates the
bitmask for LWT flag in prepared statement metadata in order to be
used for lookup in a client library. The associated bits of code in
`cql3::prepared_metadata` are adjusted to accomodate the feature.
The value for the flag is chosen on purpose to be the last bit
in the flags bitset since we don't want to possibly clash with
C* implementation in case they add more possible flag values to
prepared metadata (though there is an issue regarding that:
https://issues.apache.org/jira/browse/CASSANDRA-15746).
If it's fixed in upstream Cassandra, then we could synchronize
the value for the flag with them.
Also extend the underlying type of `flag` enum in
`cql3::prepared_metadata` to be `uint32_t` instead of `uint8_t`
because in either case flags mask is serialized as 32-bit integer.
In theory, shard-awareness extension support also should be
reworked in terms of provided minimal infrastructure, but for the
sake of simplicity, this is left to be done in a follow-up some
time later.
This solution eliminates the need to assume that all the client
drivers follow the CQL spec carefully because scylla-specific
features and protocol extensions could be enabled only in case both
server and client driver negotiate the supported feature set.
Tests: unit(dev, debug)
Signed-off-by: Pavel Solodovnikov <pa.solodovnikov@scylladb.com>
"
This series Adds a pseudo-floating-point histogram implementation.
The histogram is used for time_estimated_histogram a histogram for latency tracking and then used in storage_proxy as a more efficient with a higher resolution histogram.
Follow up series would use the new histogram in other places in the system and will add an implementation that supports lower values.
Fixes#5815Fixes#4746
"
* amnonh-quicker_estimated_histogram:
storage_proxy: use time_estimated_histogram for latencies
test/boost/estimated_histogram_test
utils/histogram_metrics_helper Adding histogram converter
utils/estimated_histogram: Adding approx_exponential_histogram
5ceb20c439 switched --enable-dpdk
to a tristate switch, but forgot that add_tristate() prepends
--enable and --disable itself; so now the switch looks like
--enable-enable-dpdk and --disable-enable-dpdk.
Fix by removing the "enable-" prefix.
"
This is part of the work for replacing global sstring variables with
constexpr std::string_view ones.
To have std::string_view values we have to convert a few APIs to take
std::string_view instead of sstring references.
The API conversions are complicated by the fact that
std::unordered_map doesn't support heterogeneous lookup, so we need
another hash map.
The one provided by abseil seems like a natural choice since it has an
API that looks like what is being proposed for c++
(http://www.open-std.org/JTC1/SC22/WG21/docs/papers/2019/p1690r0.html)
but is also much faster.
A nice side effect is that this series is a 0.46% win in
perf_simple_query --duration 16 --smp 1 -m4G
Over 500 runs with randomized section layout and environment on each
run.
"
* 'espindola/absl-v10' of https://github.com/espindola/scylla:
database: Use a flat_hash_map for _ks_cf_to_uuid
database: Use flat_hash_map for _keyspaces
Add absl wrapper headers
build: Link with abseil
cofigure: Don't overwrite seastar_cflags
Add abseil as a submodule
Using these instead of using the absl headers directly adds support
for heterogeneous lookup with sstring as key.
The is no gain from having the hash function inline, so this
implements it in a .cc file.
Signed-off-by: Rafael Ávila de Espíndola <espindola@scylladb.com>
It is a pity we have to list so many libraries, but abseil doesn't
provide a .pc file.
Signed-off-by: Rafael Ávila de Espíndola <espindola@scylladb.com>
The variable seastar_cflags was being used for flags passed to seastar
and for flags extracted from the seastar.pc file.
This introduces a new variable for the flags extracted from the
seastar.pc file.
Signed-off-by: Rafael Ávila de Espíndola <espindola@scylladb.com>
"
The cql_server and thrift are "owned" by storage_service for
the sake of managing those, i.e. starting and stopping. Since
other services (still) need the storage_service this creates
dependencies loops.
This set makes the client services independent from the storage
service. As a consequence of it the auth service is also removed
from storage_service and put standalone. This, in turn, sets
some tests free from the need to start and stop auth and makes
one step towards NOT join_cluster()-ing in unit tests.
Also the set fixes few wierd races on scylla start and stop
that can trigger local_is_initialized() asserts, and one case of
unclear aborted shutdown when client services remain running
till the scylla process exit.
Yet another benefit is localization of "isolating" functionality
that sits deeper in storage_service than it should.
One thing that's not completely clean after it is the need for cql
server to continue referencing the service_memory_limiter semaphore
from the storage_service, but this will go away with one of the
next sets.
tests: unit(debug), manual start-stop,
nodetool check of cql/thrift start/stop
"
* 'br-split-transport-1' of https://github.com/xemul/scylla:
storage_service: Isolate isolator
auth: Move away from storage_service
auth: Move start-stop code into main
main: Don't forget to stop cql/thrift when start is aborted
thrift_controller: Switch on standalone
thrift_controller: Pass one through management API
thrift_controller: Move the code into thrift/
thrift_controller: Introduce own lock for management
thrift: Wrap start/stop/is_running code into a class
cql_controller: Switch on standalone
cql_controller: Pass one through management API
cql_controller: Move the code into transport/
cql_controller: Introduce own lock for management
cql: Wrap start/stop/is_running code into a class
api: Tune reg/unreg of client services control endpoints
LCS and SCTS already have their own files, reducing the clutter in
compaction_strategy.cc. Do the same for TWCS. I am doing this in
preparation to add more functions.
Signed-off-by: Glauber Costa <glauber@scylladb.com>
Message-Id: <20200611230906.409023-6-glauber@scylladb.com>
