It now does nothing but wraps make_lw_shared<one_session_records>()
call. Callers can do it on their own thus facilitating further
list-initialization patching
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
For that to happen the value evaluation is moved from the
init_session_records() into a private trace_state helper as it checks
the props values initialized earlier
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
This object is constructed via one_session_records thus the latter needs
to pass some arguments along
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
The CQL binary protocol introduced "unset" values in version 4
of the protocol. Unset values can be bound to variables, which
cause certain CQL fragments to be skipped. For example, the
fragment `SET a = :var` will not change the value of `a` if `:var`
is bound to an unset value.
Unsets, however, are very limited in where they can appear. They
can only appear at the top-level of an expression, and any computation
done with them is invalid. For example, `SET list_column = [3, :var]`
is invalid if `:var` is bound to unset.
This causes the code to be littered with checks for unset, and there
are plenty of tests dedicated to catching unsets. However, a simpler
way is possible - prevent the infiltration of unsets at the point of
entry (when evaluating a bind variable expression), and introduce
guards to check for the few cases where unsets are allowed.
This is what this long patch does. It performs the following:
(general)
1. unset is removed from the possible values of cql3::raw_value and
cql3::raw_value_view.
(external->cql3)
2. query_options is fortified with a vector of booleans,
unset_bind_variable_vector, where each boolean corresponds to a bind
variable index and is true when it is unset.
3. To avoid churn, two compatiblity structs are introduced:
cql3::raw_value{,_view}_vector_with_unset, which can be constructed
from a std::vector<raw_value{,_view/}>, which is what most callers
have. They can also be constructed with explicit unset vectors, for
the few cases they are needed.
(cql3->variables)
4. query_options::get_value_at() now throws if the requested bind variable
is unset. This replaces all the throwing checks in expression evaluation
and statement execution, which are removed.
5. A new query_options::is_unset() is added for the users that can tolerate
unset; though it is not used directly.
6. A new cql3::unset_operation_guard class guards against unsets. It accepts
an expression, and can be queried whether an unset is present. Two
conditions are checked: the expression must be a singleton bind
variable, and at runtime it must be bound to an unset value.
7. The modification_statement operations are split into two, via two
new subclasses of cql3::operation. cql3::operation_no_unset_support
ignores unsets completely. cql3::operation_skip_if_unset checks if
an operand is unset (luckily all operations have at most one operand that
tolerates unset) and applies unset_operation_guard to it.
8. The various sites that accept expressions or operations are modified
to check for should_skip_operation(). This are the loops around
operations in update_statement and delete_statement, and the checks
for unset in attributes (LIMIT and PER PARTITION LIMIT)
(tests)
9. Many unset tests are removed. It's now impossible to enter an
unset value into the expression evaluation machinery (there's
just no unset value), so it's impossible to test for it.
10. Other unset tests now have to be invoked via bind variables,
since there's no way to create an unset cql3::expr::constant.
11. Many tests have their exception message match strings relaxed.
Since unsets are now checked very early, we don't know the context
where they happen. It would be possible to reintroduce it (by adding
a format string parameter to cql3::unset_operation_guard), but it
seems not to be worth the effort. Usage of unsets is rare, and it is
explicit (at least with the Python driver, an unset cannot be
introduced by ommission).
I tried as an alternative to wrap cql3::raw_value{,_view} (that doesn't
recognize unsets) with cql3::maybe_unset_value (that does), but that
caused huge amounts of churn, so I abandoned that in favor of the
current approach.
Closes#12517
After fcb8d040 ("treewide: use Software Package Data Exchange
(SPDX) license identifiers"), many dual-licensed files were
left with empty comments on top. Remove them to avoid visual
noise.
Closes#10562
Instead of lengthy blurbs, switch to single-line, machine-readable
standardized (https://spdx.dev) license identifiers. The Linux kernel
switched long ago, so there is strong precedent.
Three cases are handled: AGPL-only, Apache-only, and dual licensed.
For the latter case, I chose (AGPL-3.0-or-later and Apache-2.0),
reasoning that our changes are extensive enough to apply our license.
The changes we applied mechanically with a script, except to
licenses/README.md.
Closes#9937
The tracing code assumes that query_option_names and query_option_values
vectors always have the same length as the prepared_statements vector,
but it's not true. E.g. if one of the statements in a batch is
incorrect, it will create a discrepancy between the number of prepared
statements and the number of bound names and values, which currently
leads to a segmentation fault.
To overcome the problem, all three vectors are integrated into a single
vector, which makes size mismatches impossible.
Tested manually with code that triggers a failure while executing
a batch statement, because the Python driver performs driver-side
validation and thus it's hard to create a test case which triggers
the problem.
closes: #9221
The write paths in storage_proxy pass replica sets as
std::unordered_set<gms::inet_address>. This is a complex type, with
N+1 allocations for N members, so we change it to a small_vector (via
inet_address_vector_replica_set) which requires just one allocation, and
even zero when up to three replicas are used.
This change is more nuanced than the corresponding change to the read path
abe3d7d7 ("Merge 'storage_proxy: use small_vector for vectors of
inet_address' from Avi Kivity"), for two reasons:
- there is a quadratic algorithm in
abstract_write_response_handler::response(): it searches for a replica
and erases it. Since this happens for every replica, it happens N^2/2
times.
- replica sets for writes always include all datacenters, while reads
usually involve just one datacenter.
So, a write to a keyspace that has 5 datacenters will invoke 15*(15-1)/2
=105 compares.
We could remove this by sending the index of the replica in the replica
set to the replica and ask it to include the index in the response, but
I think that this is unnecessary. Those 105 compares need to be only
105/15 = 7 times cheaper than the corresponding unordered_set operation,
which they surely will. Handling a response after a cross-datacenter round
trip surely involves L3 cache misses, and a small_vector reduces these
to a minimum compared to an unordered_set with its bucket table, linked
list walking and managent, and table rehashing.
Tests using perf_simple_query --write --smp 1 --operations-per-shard 1000000
--task-quota-ms show two allocations removed (as expected) and a nice
reduction in instructions executed.
before: median 204842.54 tps ( 54.2 allocs/op, 13.2 tasks/op, 49890 insns/op)
after: median 206077.65 tps ( 52.2 allocs/op, 13.2 tasks/op, 49138 insns/op)
Closes#8847
We want to change the internals of cql3::raw_value{_view}.
However, users of cql3::raw_value and cql3::raw_value_view often
use them by extracting the internal representation, which will be different
after the planned change.
This commit prepares us for the change by making all accesses to the value
inside cql3::raw_value(_view) be done through helper methods which don't expose
the internal representation publicly.
After this commit we are free to change the internal representation of
raw_value_{view} without messing up their users.
The mechanism of session record params is currently only used
to store query strings and a couple more params like consistency level,
but since we now have more frontends than just CQL and Thrift,
it would be nice to also allow the users to put custom parameters in
there.
An immediate first user of this mechanism would be alternator,
which is going to put the operation type under the "alternator_op" key.
The operation type is not part of the query string due to how DynamoDB's
protocol works - the op type is stored separately in the HTTP header.
While it's possible to extract the operation type from the session_id,
it might not be the case once #2572 is implemented.
Currently query_options objects is passed to a trace stopping function
which makes it mandatory to make them alive until the end of the
query. The reason for that is to add prepared statement parameters to
the trace. All other query options that we want to put in the trace are
copied into trace_state::params_values, so lets copy prepared statement
parameters there too. Trace enabled case will become a little bit more
expensive but on the other hand we can drop a continuation that holds
query_options object alive from a fast path. It is safe to drop the call
to stop_foreground_prepared() here since The tracing will be stopped
in process_request_one().
Message-Id: <20191205102026.GJ9084@scylladb.com>
Replace stdx::optional and stdx::string_view with the C++ std
counterparts.
Some instances of boost::variant were also replaced with std::variant,
namely those that called seastar::visit.
Scylla now requires GCC 8 to compile.
Signed-off-by: Duarte Nunes <duarte@scylladb.com>
Message-Id: <20190108111141.5369-1-duarte@scylladb.com>
Add a new "response_size" column to system_traces.sessions and store a size of an uncompressed response
for a traced query.
Signed-off-by: Vlad Zolotarov <vladz@scylladb.com>
Add a new column "request_size" to system_traces.sessions and store
the uncompressed request frame data size.
Signed-off-by: Vlad Zolotarov <vladz@scylladb.com>
Store the prepared statement positional parameters values in the
corresponding system_traces.sessions entry in the 'parameters' column
(which has a map<text,text> type).
Parameters are stored as a pair of "param[X]" : "value", where X is
the index of the parameter starting from 0 and the "value" is the first
64 characters of the parameter's value string representation.
If parameters were given with their names attached (see the description
on bit 0x40 of QUERY flags in the CQL binary protocol specification) then
parameters are going to be stored in the "param[X](<bound variable name>)" : "value"
form.
If the value's string representation is longer than 64 characters then the "value" will
contain only first 64 characters of it and will have the "..." at
the end.
For a BATCH of prepared statements the parameter "name" will have a form of
param[Y][X] where Y is the index of the corresponding prepared statement
in the BATCH and X is the index of the parameter. Both X and Y start from
0.
Note:
Had to switch to boost::range::find() in sstables::big_sstable_set in order to
address the "ambiguous overload" compilation error.
Signed-off-by: Vlad Zolotarov <vladz@scylladb.com>
Similarly to the regular QUERY of EXECUTE we want to see the actual
queries statement that were part of the BATCH.
If a traced query has only a single statement to execute then its statement will be stored in a form 'query':'<statement>'.
If there are two or more queries (BATCH) then statements of each query in the BATCH will be stored in a form 'query[X]':'<statement>', where X is the index of the query in the
BATCH starting from 0.
Signed-off-by: Vlad Zolotarov <vladz@scylladb.com>
Hide it inside the trace_state.cc in order to avoid future circular
dependencies with other .hh files.
Signed-off-by: Vlad Zolotarov <vladz@scylladb.com>
Make state session creation, stop_forground() and tracing::trace(...) methods
noexcept.
Most of them have already been implemented the way that they won't throw
but this patch makes it official...
Signed-off-by: Vlad Zolotarov <vladz@scylladb.com>
The main idea is to log queries that take "too long" to complete.
The "too long" is above the given threshold.
To achieve the above this patch does the following:
- Introduce two new properties to the tracing::trace_state:
- "Full tracing": when the tracing of this query was explicitly requested.
In this state we will record all possible traces related to this query:
both on the coordinator and on any replica involved.
- "Log slow query": when slow query logging is enabled.
If slow query logging is enabled and a session's "duration" is above
the specified threshold we will create a record in the "slow queries log"
and write all trace records created on the coordinator and on a replica
if a replica's session lasts longer than that threshold.
(We will propagate the Coordinator's slow query logging threshold to replicas
in the context of a specific tracing/logging session).
The properties above are independent, namely they may be enabled and/or disabled
independently and any combination of them is legal (naturally, creating a tracing
session when both states above are disabled makes no sense).
- Instrument the tracing::tracing service to allow the following:
- Enable/disable slow query logging.
- Set/get the slow query duration threshold (in microseconds).
- Set/get the slow query log record TTL value (in seconds).
- Instrument the trace_keyspace_helper to write a slow query log entry
when requested.
- The slow query logging is disabled by default and the threshold is set to half a second.
- The TTL of a slow log record is set to 86400 seconds by default.
- It makes sense to use the same "slow query logging threshold" and a "slow query record TTL"
both on a coordinator and on a replica Nodes in a context of the same tracing session:
- Pass both TTL and a threshold to the replica in a trace_info.
This patch also implements the new slow query logging specific logic:
- Don't write the pending tracing records before the end of a tracing session
until "duration" reaches the logging threshold.
- Don't build the parameters<sstring, sstring> map unless we know we will write it
to I/O.
Signed-off-by: Vlad Zolotarov <vladz@cloudius-systems.com>
- Instead of keeping separate booleans introduce a trace_state_props_set enum_set and
pass it around instead of separate booleans.
- Change the trace_info to hold this value in addition to write_on_close. Initialize
a corresponding bit in an enum_set based on a write_on_close value in a trace_info
constructor for a backward compatibility.
- Separate a trace_state constructor into two:
- For a primary session object.
- For a secondary session object.
Signed-off-by: Vlad Zolotarov <vladz@cloudius-systems.com>
The condition in question is sanity check for a SW bug.
This SW bug (if occurs) is not critical - there is an additional protection
against it in the stop_foreground_and_write().
Having said all that, since we shell not throw from a destructor,
replace throwing of a std::logic_error with an logger error message.
Signed-off-by: Vlad Zolotarov <vladz@cloudius-systems.com>
Message-Id: <1471773320-7398-1-git-send-email-vladz@cloudius-systems.com>
A session's "duration" should be a time it took to
handle a request, which is a time till response to a user.
In other words - till a consistency level is reached.
Before this patch is was a time that takes a complete
handling of a request, which is the time it takes to handle
all replicas and not only those required to reach a CL.
This patch fixes this situation by extending the trace_state's state
values to 3 states: inactive, foreground and background.
A primary session may be in 3 states:
- "inactive": between the creation and a begin() call.
- "foreground": after a begin() call and before a
stop_foreground_and_write() call.
- "background": after a stop_foreground_and_write() call and till the
state object is destroyed.
- Traces are not allowed while state is in an "inactive" state.
- The time the primary session was in a "foreground" state is the time
reported as a session's "duration".
- Traces that have arrived during the "background" state will be recorded
as usual but their "elapsed" time will be greater or equal to the
session's "duration".
Secondary sessions may only be in an "inactive" or in a "foreground"
states.
Signed-off-by: Vlad Zolotarov <vladz@cloudius-systems.com>
Use a per-shard tracing records budget instead
of maintaining a fixed-size per-session records budget and
a per-shard sessions budget.
The original policy could lead to some irrational situations,
when we have a single tracing session that creates a substantial
amount of records that we can handle but we would start dropping
new records after it surpasses the per-session limit.
The new policy handles a per-shard trace records budget that is
being consumed by each trace() call and by a primary session destructor
when a session record is created.
Each active record may only be in one of the following states:
- cached: stored in its session's object. When record is in this state
it's not going to be written to I/O during the next write event.
- pending for write: when record is in this state it's going to be written
to I/O during the next write event.
- flushing: the record is being currently written to the I/O.
There are counters of the total amount of records in each state above.
Each record may only be in a specific state at every point of time and
thereby it must be accounted only in one and only one of the three
counters.
The sum of all three counters should not be greater than
(max_pending_trace_records + write_event_records_threshold) at any time
(actually it can get as high as a value above plus (max_pending_sessions)
if all sessions are primary but we won't take this into an account for
simplicity).
The same is about the number of outstanding sessions: it may not be greater
than (max_pending_sessions + write_event_sessions_threshold) at any time.
If total number of tracing records is greater or equal to the limit
above, the new trace point is going to be dropped.
If current number or records plus the expected number of trace records
per session (exp_trace_events_per_session) is greater than the limit
above new sessions will be dropped. A new session will also be dropped if
there are too many active sessions.
When the record or a session is dropped the appropriate statistics
counters are updated and there is a rate-limited warning message printed
to the log.
Every time a number of records pending for write is greater or equal to
(write_event_records_threshold) or a number of sessions pending for
write is greater or equal to (write_event_sessions_threshold) a write
event is issued.
Every 2 seconds a timer would write all pending for write records
available so far.
Signed-off-by: Vlad Zolotarov <vladz@cloudius-systems.com>
Before this patch the interaction between the layers above was as follows:
- trace_state was passing the trace event data to a backend object every
time trace() method was called.
- trace_state was passing the session data to a backend object in a destructor.
- A backend object was storing this data in a form of lambda where all data
above was caught in a capture list. This was primarily done in order to
delay the call for make_xxx_mutation(). Lambdas were stored in a map by a session
ID and they were executed when a kick() method was called.
- A tracing::tracing object was periodically calling a kick() method of a
backend that was initiating a write of all pending data to the storage.
All backend methods used in the described above interactions were virtual.
Thereby, for instance, for each and every trace record we were calling a virtual method that was
receiving a significant amount of parameters, store a lambda in a map and return.
This is clearly a suboptimal way of using virtual functions since we prevent a compiler
from inlining an obviously inlinable operations.
This patch changes the interaction scheme to be as follows:
- Trace events and session data are stored and passed around in a form of structs
that hold all relevant information (no more lambdas).
- As long as a trace session is active its data is aggregated inside the corresponding
trace_state object.
- The object containing all records is passed and stored as a lw_shared_ptr to save extra
copies and to shorten capture lists.
- All aggregated data is passed to a tracing::tracing object in a trace_state destructor.
The data is stored in a std::deque in a tracing::tracing object (instead of a map by a session ID).
- A single backend's virtual method call writes all data aggregated so far (kick()
method is not needed any more), every time a write event occurs.
- Backend has only one virtual method now:
- Write a bulk of sessions' data aggregated so far.
- Backend's virtual method receives a records bulk object by reference.
As a result:
- A latency of a single trace event that has no formatting improved from 0.2us to 0.1us.
Signed-off-by: Vlad Zolotarov <vladz@cloudius-systems.com>
Sometimes we want to be able to set "params" map after we
started a tracing session, e.g. when the parameters values,
like a consistency level parsed from the "options" part of a binary frame,
are available only after some heavy part of a flow we would like
to trace.
This patch includes the following changes:
- No longer pass a map to the begin().
- Limit the parameters to the known set.
- Define a method to set each such parameter and save its
value till the final sstring->sstring map is created.
- Construct the final sstring->sstring map in the destructor of the trace_state
object in order to defer all the formatting to be after the traced flow.
Signed-off-by: Vlad Zolotarov <vladz@cloudius-systems.com>
Add an support for passing a format string plus positional parameters
for creation of a trace point message.
Format string should be given in a fmt library native format described
here: http://fmtlib.net/latest/syntax.html#syntax .
Signed-off-by: Vlad Zolotarov <vladz@cloudius-systems.com>
Prevent the destruction of tracing::tracing instances while there
are still tracing::trace_state objects that are using it:
- Make tracing::tracing inherit from seastar::async_sharded_service<tracing::tracing>.
- Grab a tracing::tracing.shared_from_this() in each
tracing::trace_state object using it.
- Use a saved pointer to the local tracing::tracing instance in a destructor
instead of accessing it via tracing::get_local_tracing_instance()
to avoid "local is not initialized" assert when sessions are
being destroyed after the service was stopped.
Signed-off-by: Vlad Zolotarov <vladz@cloudius-systems.com>
In names of functions and variables:
s/flush_/write_/
s/store_/write_/
In a i_tracing_backend_helper:
s/flush()/kick()/
Signed-off-by: Vlad Zolotarov <vladz@cloudius-systems.com>
A tracing session life cycle includes 3 stages:
1) Active: when new trace records are being added to this session.
2) Pending for flushing to a storage: when session is over but not
yet flushed to the storage ("backend").
3) Flushing: when session's records are being flushed to the storage
and this process is not yet completed.
Sessions may accumulate in each of the stages above and we should limit
the maximum amount of sessions being accumulated in each of them in order to avoid OOM
situation.
Current in-tree implementation only limits the number of tracing sessions
accumulated in the first ("Active") stage.
Since currently every closing session is being immediately flushed (as long
as "settraceprobability" is not implemented) the second stage never accumulates
tracing sessions.
The third stage is currently not controlled at all and if, for instance, we
succeed to push enough tracing session towards a slow storage backend, they may
accumulate there consuming an uncontrolled amount of memory and may eventually consume
all of it.
This patch fixes this unpleasant situation by implying the following strategy:
- Limit the total amount of accumulated tracing sessions in all stages above together
by a static value - 2 times "flush threshold". "2 times" is needed to allow new
tracing sessions to accumulate in the stage 2 while sessions in the stage 3 are still
being processed.
- Forcefully flush sessions in the stage 2 to the storage when their count reaches a "flush
threshold".
This would ensure that there will not more than totally (2 * "flush threshold") sessions (in any stage)
on each shard.
An advantage of this strategy is its simplicity - we only need a single threshold to control all stages.
If we feel that we needed a finer graining for each stage we may add separate limits for each of them
in the future.
Signed-off-by: Vlad Zolotarov <vladz@cloudius-systems.com>
trace_state: Is a single tracing session.
tracing: A sharded service that contains an i_trace_backend_helper instance
and is a "factory" of trace_state objects.
trace_state main interface functions are:
- begin(): Start time counting (should be used via tracing::begin() wrapper).
- trace(): Create a tracing event - it's coupled with a time passed since begin()
(should be used via tracing::trace() wrapper).
- ~trace_state(): Destructor will close the tracing session.
"tracing" service main interface function is:
- start(): Initialize a backend.
- stop(): Shut down a backend.
- create_session(): Creates a new tracing session.
(tracing::end_session(): Is called by a trace_state destructor).
When trace_state needs to store a tracing event it uses a backend helper from
a "tracing" service.
A "tracing" service limits a number of opened tracing session by a static number.
If this number is reached - next sessions will be dropped.
trace_state implements a similar strategy in regard to tracing events per singe
session.
Signed-off-by: Vlad Zolotarov <vladz@cloudius-systems.com>
