1. Fix LoadTask, ChunkAllocator, TabletMeta, Brpc, the accuracy of memory track.
2. Modified some MemTracker names, deleted some unnecessary trackers, and improved readability.
3. More powerful MemTracker debugging capabilities.
4. Avoid creating TabletColumn temporary objects and improve BE startup time by 8%.
5. Fix some other details.
1. fix track bthread
- Bthread, a high performance M:N thread library used by brpc. In Doris, a brpc server response runs on one bthread, possibly on multiple pthreads. Currently, MemTracker consumption relies on pthread local variables (TLS).
- This caused pthread TLS MemTracker confusion when switching pthread TLS MemTracker in brpc server response. So replacing pthread TLS with bthread TLS in the brpc server response saves the MemTracker.
Ref: 731730da85/docs/en/server.md (bthread-local)
2. fix track vectorized query
- Added track mmap. Currently, mmap allocates memory in many places of the vectorized execution engine.
- Refactored ThreadContext to avoid dependency conflicts and make it easier to debug.
- Fix some bugs.
The patch fixes two problems.
1. Memory order problem accessing _last_patch_processed_finished and in_flight, actually _last_patch_processed_finished is redundant, so the patch removes it.
2. synchronization in join on cid.
Fix for #8725.
Early Design Documentation: https://shimo.im/docs/DT6JXDRkdTvdyV3G
Implement a new way of memory statistics based on TCMalloc New/Delete Hook,
MemTracker and TLS, and it is expected that all memory new/delete/malloc/free
of the BE process can be counted.
1.
The methods in the IndexChannel are called back in the RpcClosure in the NodeChannel.
However, this callback may occur after the whole task is finished (e.g. due to network latency),
and by that time the IndexChannel may have been destructured, so we should not call
the IndexChannel methods anymore, otherwise the BE will crash.
Therefore, we use the `_is_closed` variable and `_closed_lock` to ensure that the RPC callback function
will not call the IndexChannel's method after the NodeChannel is closed.
2.
Do not add IndexChannel to the ObjectPool.
Because when deconstruct IndexChannel, it may call the deconstruction of NodeChannel.
And the deconstruction of NodeChannel maybe time consuming(wait rpc finished).
But the ObjectPool will hold a SpinLock to destroy the objects, so it may cause CPU busy.
Due to unlimited queue in OlapScanNode and NodeChannel, memory usage can be
very large for reading and writing large table, e.g 'insert into tableB select * from tableA'.
In some scenarios, users cannot find a suitable hash key to avoid data skew, so we need to provide an additional data distribution for olap table to avoid data skew
example:
CREATE TABLE random_table
(
siteid INT DEFAULT '10',
citycode SMALLINT,
username VARCHAR(32) DEFAULT '',
pv BIGINT SUM DEFAULT '0'
)
AGGREGATE KEY(siteid, citycode, username)
DISTRIBUTED BY random BUCKETS 10
PROPERTIES("replication_num" = "1");
Co-authored-by: caiconghui1 <caiconghui1@jd.com>
1. Fix the problem of BE crash caused by destruct sequence. (close#8058)
2. Add a new BE config `compaction_task_num_per_fast_disk`
This config specify the max concurrent compaction task num on fast disk(typically .SSD).
So that for high speed disk, we can execute more compaction task at same time,
to compact the data as soon as possible
3. Avoid frequent selection of unqualified tablet to perform compaction.
4. Modify some log level to reduce the log size of BE.
5. Modify some clone logic to handle error correctly.
1. set both `tuple_offsets` and `new_tuple_offsets` in PRowBatch for compatibility
2. set FE config `repair_slow_replica` default to false
Avoid impacting the load process after upgrading.
Eg, if there are only 2 replicas, one is with high version count. After upgrade,
that replica will be set to bad, so that the load process will be stopped
because only 1 replica is alive.
3. Fix a bug that NodeChannel may be blocked at `close_wait()`
Forget to set `add_batch_finish` flag after the last rpc finished.
4. Fix a NPE of RoutineLoadScheduler
This PR mainly changes:
1. Fix bug when enable `transfer_data_by_brpc_attachment`
In `data_stream_sender`, we will send a serialized PRowBatch data to multiple Channels.
And if `transfer_data_by_brpc_attachment` is enabled, we will mistakenly clear the data in PRowBatch
after sending PRowBatch to the first Channel.
As a result, the following Channel cannot receive the correct data, causing an error.
So I use a separate buffer instead of `tuple_data` in PRowBatch to store the serialized data
and reuse it in multiple channels.
2. Fix bug that the the offset in serialized row batch may overflow
Use int64 to replace int32 offset. And for compatibility, add a new field `new_tuple_offsets` in PRowBatch.
Currently, if we encounter a problem with a replica of a tablet during the load process,
such as a write error, rpc error, -235, etc., it will cause the entire load job to fail,
which results in a significant reduction in Doris' fault tolerance.
This PR mainly changes:
1. refined the judgment of failed replicas in the load process, so that the failure of a few replicas will not affect the normal completion of the load job.
2. fix a bug introduced from #7754 that may cause BE coredump
This PR mainly changes:
1. Help to Cancel the load job ASAP when encounter unqualified data.
Solution is described in #6318 .
Also replace some std::stringstream with fmt::memory_buffer to avoid performance issues.
2. fix a NPE bug when create user with empty host
3. fix compile warning after rebasing the master(vectorization)
# Proposed changes
Issue Number: close#6238
Co-authored-by: HappenLee <happenlee@hotmail.com>
Co-authored-by: stdpain <34912776+stdpain@users.noreply.github.com>
Co-authored-by: Zhengguo Yang <yangzhgg@gmail.com>
Co-authored-by: wangbo <506340561@qq.com>
Co-authored-by: emmymiao87 <522274284@qq.com>
Co-authored-by: Pxl <952130278@qq.com>
Co-authored-by: zhangstar333 <87313068+zhangstar333@users.noreply.github.com>
Co-authored-by: thinker <zchw100@qq.com>
Co-authored-by: Zeno Yang <1521564989@qq.com>
Co-authored-by: Wang Shuo <wangshuo128@gmail.com>
Co-authored-by: zhoubintao <35688959+zbtzbtzbt@users.noreply.github.com>
Co-authored-by: Gabriel <gabrielleebuaa@gmail.com>
Co-authored-by: xinghuayu007 <1450306854@qq.com>
Co-authored-by: weizuo93 <weizuo@apache.org>
Co-authored-by: yiguolei <guoleiyi@tencent.com>
Co-authored-by: anneji-dev <85534151+anneji-dev@users.noreply.github.com>
Co-authored-by: awakeljw <993007281@qq.com>
Co-authored-by: taberylyang <95272637+taberylyang@users.noreply.github.com>
Co-authored-by: Cui Kaifeng <48012748+azurenake@users.noreply.github.com>
## Problem Summary:
### 1. Some code from clickhouse
**ClickHouse is an excellent implementation of the vectorized execution engine database,
so here we have referenced and learned a lot from its excellent implementation in terms of
data structure and function implementation.
We are based on ClickHouse v19.16.2.2 and would like to thank the ClickHouse community and developers.**
The following comment has been added to the code from Clickhouse, eg:
// This file is copied from
// https://github.com/ClickHouse/ClickHouse/blob/master/src/Interpreters/AggregationCommon.h
// and modified by Doris
### 2. Support exec node and query:
* vaggregation_node
* vanalytic_eval_node
* vassert_num_rows_node
* vblocking_join_node
* vcross_join_node
* vempty_set_node
* ves_http_scan_node
* vexcept_node
* vexchange_node
* vintersect_node
* vmysql_scan_node
* vodbc_scan_node
* volap_scan_node
* vrepeat_node
* vschema_scan_node
* vselect_node
* vset_operation_node
* vsort_node
* vunion_node
* vhash_join_node
You can run exec engine of SSB/TPCH and 70% TPCDS stand query test set.
### 3. Data Model
Vec Exec Engine Support **Dup/Agg/Unq** table, Support Block Reader Vectorized.
Segment Vec is working in process.
### 4. How to use
1. Set the environment variable `set enable_vectorized_engine = true; `(required)
2. Set the environment variable `set batch_size = 4096; ` (recommended)
### 5. Some diff from origin exec engine
https://github.com/doris-vectorized/doris-vectorized/issues/294
## Checklist(Required)
1. Does it affect the original behavior: (No)
2. Has unit tests been added: (Yes)
3. Has document been added or modified: (No)
4. Does it need to update dependencies: (No)
5. Are there any changes that cannot be rolled back: (Yes)
If an load task has a relatively short timeout, then we need to ensure that
each RPC of this task does not get blocked for a long time.
And an RPC is usually blocked for two reasons.
1. handling "memory exceeds limit" in the RPC
If the system finds that the memory occupied by the load exceeds the threshold,
it will select the load channel that occupies the most memory and flush the memtable in it.
this operation is done in the RPC, which may be more time consuming.
2. close the load channel
When the load channel receives the last batch, it will end the task.
It will wait for all memtables flushes to finish synchronously. This process is also time consuming.
Therefore, this PR solves this problem by.
1. Use timeout to determine whether it is a high-priority load task
If the timeout of an load task is relatively short, then we mark it as a high-priority task.
2. not processing "memory exceeds limit" for high priority tasks
3. use a separate flush thread to flush memtable for high priority tasks.
Added bprc stub cache check and reset api, used to test whether the bprc stub cache is available, and reset the bprc stub cache
add a config used for auto check and reset bprc stub
This CL mainly changes:
1. Add star schema benchmark tools in `tools/ssb-tools`, for user to easy load and test with SSB data set.
2. Disable the segment cache for some read scenario such as compaction and alter operation.(Fix#6924 )
3. Fix a bug that `max_segment_num_per_rowset` won't work(Fix#6926)
4. Enable `enable_batch_delete_by_default` by default.
1. Fix NPE in ReplicasProcNode when backend does not exist
2. Forbid the create table like statement to specify the view.
3. Check self ip when starting FE to see if it use the origin ip.
4. Modify the error msg of tablet sink to show more detail errors.
In the previous broker load, multiple OlapTableSinks would send data to the same LoadChannel,
and because of the lock granularity problem, LoadChannel could only process these requests serially,
which made it impossible to make full use of cluster resources.
This CL modifies the related locks so that LoadChannel can process these requests in parallel.
In the test, with a size of 20G, the load speed of 334 million rows of data in 3 nodes has been
increased from 9min to 5min, and after enabling 2 concurrency, it can be increased to 3min.
Also modify the profile of load job.
ThreadSanitizer, aka TSAN, is a useful tool to detect multi-thread
problems, such as data race, mutex problems, etc.
We should detect TSAN problems for Doris BE, both unit tests and
server should pass through TSAN mode, to make Doris more robustness.
This is the very beginning patch to fix TSAN problems, and some
difficult problems are suppressed in file 'tsan_suppressions', you
can suppress these problems by setting:
export TSAN_OPTIONS="suppressions=tsan_suppressions"
before running:
`BUILD_TYPE=tsan ./run-be-ut.sh --run`
The close method of OlapTabletSink may be called twice.
In the open_internal() method of plan_fragment_executor, close is called once.
If an error occurs in this call, it will be called again in fragment_mgr.
So here we use a flag to prevent repeated close operations.
Co-authored-by: morningman <chenmingyu@baidu.com>
BE can not graceful exit because some threads are running in endless
loop. This patch do the following optimization:
- Use the well encapsulated Thread and ThreadPool instead of std::thread
and std::vector<std::thread>
- Use CountDownLatch in thread's loop condition to avoid endless loop
- Introduce a new class Daemon for daemon works, like tcmalloc_gc,
memory_maintenance and calculate_metrics
- Decouple statistics type TaskWorkerPool and StorageEngine notification
by submit tasks to TaskWorkerPool's queue
- Reorder objects' stop and deconstruct in main(), i.e. stop network
services at first, then internal services
- Use libevent in pthreads mode, by calling evthread_use_pthreads(),
then EvHttpServer can exit gracefully in multi-threads
- Call brpc::Server's Stop() and ClearServices() explicitly
We make all MemTrackers shared, in order to show MemTracker real-time consumptions on the web.
As follows:
1. nearly all MemTracker raw ptr -> shared_ptr
2. Use CreateTracker() to create new MemTracker(in order to add itself to its parent)
3. RowBatch & MemPool still use raw ptrs of MemTracker, it's easy to ensure RowBatch & MemPool destructor exec
before MemTracker's destructor. So we don't change these code.
4. MemTracker can use RuntimeProfile's counter to calc consumption. So RuntimeProfile's counter need to be shared
too. We add a shared counter pool to store the shared counter, don't change other counters of RuntimeProfile.
Note that, this PR doesn't change the MemTracker tree structure. So there still have some orphan trackers, e.g. RowBlockV2's MemTracker. If you find some shared MemTrackers are little memory consumption & too time-consuming, you could make them be the orphan, then it's fine to use the raw ptr.
ImplementaItion Notes
NodeChannel
_cur_batch -> _pending_batches: when _cur_batch is filled up, move it to _pending_batches.
add_row() just produce batches.
try_send_and_fetch_status() tries to consume one pending batch. If has in flight packet, skip send in this round.
So we can add one sender thread to be in charge of all node channels try_send.
IndexChannel
init(), open() stay the same.
Use for_each_node_channel() to expose the detailed changes of NodeChannel.(It's more easy to read & modify)
Sender thread
See func OlapTableSink::_send_batch_process()
Why use polling？
If we use wait/notify, it will notify when generate a new batch. We can't skip sending this batch, coz it won't notify the same batch again. So wait/notify can't avoid blocking simply.
So I choose polling.
It's wasting to continuously try_send(), but it's difficult to set the suitable polling interval. Thus, I add std::this_thread::yield() to give up the time slice, give priority to other process/threads (if there are other process/threads waiting in the queue).
[Load]
When performing a long-time load job, the following errors may occur. Causes the load to fail.
load channel manager add batch with unknown load id: xxx
There is a case of this error because Doris opened an unrelated channel during the load
process. This channel will not receive any data during the entire load process. Therefore,
after a fixed timeout, the channel will be released.
And after the entire load job is completed, it will try to close all open channels. When it try to
close this channel, it will find that the channel no longer exists and an error is reported.
This CL will pass the timeout of load job to the load channel, so that the timeout of load channels
will be same as load job's.
Use same UUID as query ID and load ID of a load execution plan.
Each load execution plan has a load ID, and as a plan, there is also a query ID.
We can use same UUID as query ID and load ID, for tracing the load process more easily.
Change the load ID when retrying a load execution plan.
When a load execution plan retry, the load ID should be changed, otherwise BE can not
distinguish the old and new load requests.
Cancel the running loading task when cancelling the broker load.
When user cancel a broker load, the running loading task should also be cancelled, or
it may occupies the worker thread for a long time.
Remove the unnecessary query report when doing load execution plan.
Only the last query report is needed.
Add a new BE config tablet_writer_rpc_timeout_sec.
It is used for RPC of tablet sink. The default is 600 seconds. which is long enough for flushing
about 6GB data. The long timeout config will reduce the possibility of encountering fail to send batch error when loading.
Use streaming_load_max_mb instead of mini_load_max_mb in BE config.
Add more logs for tracing a broker load process easily.
The Operator wants to known when the job being scheduled as PENDING
and LOADING. And how long it takes to finish these sub states.
Also add 2 metrics on BE to monitor the memtable's flush time.
`memtable_flush_total` and `memtable_flush_duration_us`
Mini load is now using stream load framework. But we should keep the
mini load return behavior and result json format be same as old.
So PUBLISH_TIMEOUT error should be treated as OK in mini load.
Also add 2 counters for OlapTableSink profile:
SerializeBatchTime: time of serializing all row batch.
WaitInFlightPacketTime: time of waiting last send packet
NOTE: This patch would modify all Backend's data.
And this will cause a very long time to restart be.
So if you want to interferer your product environment,
you should upgrade backend one by one.
1. Refactoring be is to clarify the structure the codes.
2. Use unique id to indicate a rowset.
Nameing rowset with tablet_id and version will lead to
many conflicts among compaction, clone, restore.
3. Extract an rowset interface to encapsulate rowsets
with different format.
