- This commit restores statements "SYSTEM RELOAD MODEL(S)" which provide
a mechanism to update a model explicitly. It also saves potentially
unnecessary reloads of a model from disk after it's initial load.
To keep the complexity low, the semantics of "SYSTEM RELOAD MODEL(S)
was changed from eager to lazy. This means that both statements
previously immedately reloaded the specified/all models, whereas now
the statements only trigger an unload and the first call to
catboostEvaluate() does the actual load.
- Monitoring view SYSTEM.MODELS is also restored but with some obsolete
fields removed. The view was not documented in the past and for now it
remains undocumented. The commit is thus not considered a breach of
ClickHouse's public interface.
This commit moves the catboost model evaluation out of the server
process into the library-bridge binary. This serves two goals: On the
one hand, crashes / memory corruptions of the catboost library no longer
affect the server. On the other hand, we can forbid loading dynamic
libraries in the server (catboost was the last consumer of this
functionality), thus improving security.
SQL syntax:
SELECT
catboostEvaluate('/path/to/model.bin', FEAT_1, ..., FEAT_N) > 0 AS prediction,
ACTION AS target
FROM amazon_train
LIMIT 10
Required configuration:
<catboost_lib_path>/path/to/libcatboostmodel.so</catboost_lib_path>
*** Implementation Details ***
The internal protocol between the server and the library-bridge is
simple:
- HTTP GET on path "/extdict_ping":
A ping, used during the handshake to check if the library-bridge runs.
- HTTP POST on path "extdict_request"
(1) Send a "catboost_GetTreeCount" request from the server to the
bridge, containing a library path (e.g /home/user/libcatboost.so) and
a model path (e.g. /home/user/model.bin). Rirst, this unloads the
catboost library handler associated to the model path (if it was
loaded), then loads the catboost library handler associated to the
model path, then executes GetTreeCount() on the library handler and
finally sends the result back to the server. Step (1) is called once
by the server from FunctionCatBoostEvaluate::getReturnTypeImpl(). The
library path handler is unloaded in the beginning because it contains
state which may no longer be valid if the user runs
catboost("/path/to/model.bin", ...) more than once and if "model.bin"
was updated in between.
(2) Send "catboost_Evaluate" from the server to the bridge, containing
the model path and the features to run the interference on. Step (2)
is called multiple times (once per chunk) by the server from function
FunctionCatBoostEvaluate::executeImpl(). The library handler for the
given model path is expected to be already loaded by Step (1).
Fixes#27870
This commit moves the catboost model evaluation out of the server
process into the library-bridge binary. This serves two goals: On the
one hand, crashes / memory corruptions of the catboost library no longer
affect the server. On the other hand, we can forbid loading dynamic
libraries in the server (catboost was the last consumer of this
functionality), thus improving security.
SQL syntax:
SELECT
catboostEvaluate('/path/to/model.bin', FEAT_1, ..., FEAT_N) > 0 AS prediction,
ACTION AS target
FROM amazon_train
LIMIT 10
Required configuration:
<catboost_lib_path>/path/to/libcatboostmodel.so</catboost_lib_path>
*** Implementation Details ***
The internal protocol between the server and the library-bridge is
simple:
- HTTP GET on path "/extdict_ping":
A ping, used during the handshake to check if the library-bridge runs.
- HTTP POST on path "extdict_request"
(1) Send a "catboost_GetTreeCount" request from the server to the
bridge, containing a library path (e.g /home/user/libcatboost.so) and
a model path (e.g. /home/user/model.bin). Rirst, this unloads the
catboost library handler associated to the model path (if it was
loaded), then loads the catboost library handler associated to the
model path, then executes GetTreeCount() on the library handler and
finally sends the result back to the server. Step (1) is called once
by the server from FunctionCatBoostEvaluate::getReturnTypeImpl(). The
library path handler is unloaded in the beginning because it contains
state which may no longer be valid if the user runs
catboost("/path/to/model.bin", ...) more than once and if "model.bin"
was updated in between.
(2) Send "catboost_Evaluate" from the server to the bridge, containing
the model path and the features to run the interference on. Step (2)
is called multiple times (once per chunk) by the server from function
FunctionCatBoostEvaluate::executeImpl(). The library handler for the
given model path is expected to be already loaded by Step (1).
Fixes#27870
- In general, it is expected that clickhouse-*-bridges and
clickhouse-server were build from the same source version (e.g. are
upgraded "atomically"). If that is not the case, we should at least
be able to detect the mismatch and abort.
- This commit adds a URL parameter "version", defined in a header shared
by the server and bridges. The bridge returns an error in case of
mismatch.
- The version is *not* send and checked for "ping" requests (used for
handshake), only for regular requests send after handshake. This is
because the internally thrown server-side exception due to HTTP
failure does not propagate the exact HTTP error (it only stores the
error as text), and as a result, the server-side handshake code
simply retries in case of error with exponential backoff and finally
fails with a "timeout error". This is reasonable as pings typically
fail due to time out. However, without a rework of HTTP exceptions,
version mismatch during ping would also appear as "timeout" which is
too misleading. The behavior may be changed later if needed.
- Note that introducing a version parameter does not represent a
protocol upgrade itself. Bridges older than the server will simply
ignore the field. Only servers older than the bridges receive an error
but such a situation should never occur in practice.
Wall of text, sorry, but I also had to document some stuff for myself:
There are three ways to communicate data using HTTP:
- the HTTP verb: for our purposes, PUT and GET,
- the HTTP path: '/ping', '/request' etc.,
- the HTTP URL parameter(s), e.g. 'method=libNew&dictionary_id=1234'
The bridge will use different handlers for communication with the
external dictionary library and for communication with the catboost
library. Handlers are created based on a combination of the HTTP verb
and the HTTP method. More specifically, there will be combinations
- GET + '/extdict_ping'
- PUT + '/extdict_request'
- GET + '/catboost_ping'
- PUT + '/catboost_request'.
For each combination, the bridge expects a certain set of URL
parameters, e.g. for the first combination parameter "dictionary_id" is
expected.
Starting with this commit, the library-bridge creates handlers based on
the first two combinations (the latter two combinations will be added
later). This makes the handler creation mechanism consistent with it's
counterpart in xdbc-bridge.
For that, it was necessary to make both IBridgeHelper methods
"getMainURI()" and "getPingURI()" pure virtual so that derived classes
(LibraryBridgeHelper and XDBCBridgeHelper) must provide custom URLs with
custom paths.
Side note 1: Previously, LibraryBridgeHelper sent HTTP URL parameter
"method=ping" during handshake (PING) but the library-bridge ignored
that parameter. We now omit this parameter, i.e.
LibraryBridgeHelper::PING was removed. Again, this makes things
consistent with xdbc-bridge.
Side note 2: xdbc-bridge is unchanged in this commit. Therefore,
XDBCBridgeHelper now uses the HTTP paths previously in the base class.
For funny reason, XDBCBridgeHelper did not use
IBridgeHelper::getMainURI() - it generates the URLs by itself. I kept it
that way for now but provided an implementation of getMainURI() anyways.
- Rename generic file and identifier names in library-bridge to
something more dictionary-specific. This is needed because later on,
catboost will be integrated into library-bridge.
- Also: Some smaller fixes like typos and un-inlining non-performance
critical code.
- The logic remains unchanged in this commit.
A simple HelloWorld program with zero includes except iostream triggers
a build of ca. 2000 source files. The reason is that ClickHouse's
top-level CMakeLists.txt overrides "add_executable()" to link all
binaries against "clickhouse_new_delete". This links against
"clickhouse_common_io", which in turn has lots of 3rd party library
dependencies ... Without linking "clickhouse_new_delete", the number of
compiled files for "HelloWorld" goes down to ca. 70.
As an example, the self-extracting-executable needs none of its current
dependencies but other programs may also benefit.
In order to restore access to the original "add_executable()", the
overriding version is now prefixed. There is precedence for a
"clickhouse_" prefix (as opposed to "ch_"), for example
"clickhouse_split_debug_symbols". In general prefixing makes sense also
because overriding CMake commands relies on undocumented behavior and is
considered not-so-great practice (*).
(*) https://crascit.com/2018/09/14/do-not-redefine-cmake-commands/
- TSA is a static analyzer build by Google which finds race conditions
and deadlocks at compile time.
- It works by associating a shared member variable with a
synchronization primitive that protects it. The compiler can then
check at each access if proper locking happened before. A good
introduction are [0] and [1].
- TSA requires some help by the programmer via annotations. Luckily,
LLVM's libcxx already has annotations for std::mutex, std::lock_guard,
std::shared_mutex and std::scoped_lock. This commit enables them
(--> contrib/libcxx-cmake/CMakeLists.txt).
- Further, this commit adds convenience macros for the low-level
annotations for use in ClickHouse (--> base/defines.h). For
demonstration, they are leveraged in a few places.
- As we compile with "-Wall -Wextra -Weverything", the required compiler
flag "-Wthread-safety-analysis" was already enabled. Negative checks
are an experimental feature of TSA and disabled
(--> cmake/warnings.cmake). Compile times did not increase noticeably.
- TSA is used in a few places with simple locking. I tried TSA also
where locking is more complex. The problem was usually that it is
unclear which data is protected by which lock :-(. But there was
definitely some weird code where locking looked broken. So there is
some potential to find bugs.
*** Limitations of TSA besides the ones listed in [1]:
- The programmer needs to know which lock protects which piece of shared
data. This is not always easy for large classes.
- Two synchronization primitives used in ClickHouse are not annotated in
libcxx:
(1) std::unique_lock: A releaseable lock handle often together with
std::condition_variable, e.g. in solve producer-consumer problems.
(2) std::recursive_mutex: A re-entrant mutex variant. Its usage can be
considered a design flaw + typically it is slower than a standard
mutex. In this commit, one std::recursive_mutex was converted to
std::mutex and annotated with TSA.
- For free-standing functions (e.g. helper functions) which are passed
shared data members, it can be tricky to specify the associated lock.
This is because the annotations use the normal C++ rules for symbol
resolution.
[0] https://clang.llvm.org/docs/ThreadSafetyAnalysis.html
[1] https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/42958.pdf
Here is oneliner:
$ gg 'LOG_\(DEBUG\|TRACE\|INFO\|TEST\|WARNING\|ERROR\|FATAL\)([^,]*, [a-zA-Z]' -- :*.cpp :*.h | cut -d: -f1 | sort -u | xargs -r sed -E -i 's#(LOG_[A-Z]*)\(([^,]*), ([A-Za-z][^,)]*)#\1(\2, fmt::runtime(\3)#'
Note, that I tried to do this with coccinelle (tool for semantic
patchin), but it cannot parse C++:
$ cat fmt.cocci
@@
expression log;
expression var;
@@
-LOG_DEBUG(log, var)
+LOG_DEBUG(log, fmt::runtime(var))
I've also tried to use some macros/templates magic to do this implicitly
in logger_useful.h, but I failed to do so, and apparently it is not
possible for now.
Signed-off-by: Azat Khuzhin <a.khuzhin@semrush.com>
v2: manual fixes
Signed-off-by: Azat Khuzhin <a.khuzhin@semrush.com>
