mirror of
https://github.com/csunny/DB-GPT.git
synced 2025-09-11 13:58:58 +00:00
update
This commit is contained in:
csunny
188
pilot/app.py
188
pilot/app.py
@@ -33,7 +33,193 @@ def knowledged_qa_demo(text_list):
|
|||||||
|
|
||||||
|
|
||||||
def get_answer(q):
|
def get_answer(q):
|
||||||
base_knowledge = """ 这是一段测试文字 """
|
base_knowledge = """
|
||||||
|
执行计划是对一条 SQL 查询语句在数据库中执行过程的描述。用户可以通过 EXPLAIN 命令查看优化器针对指定 SQL 生成的逻辑执行计划。
|
||||||
|
|
||||||
|
如果要分析某条 SQL 的性能问题，通常需要先查看 SQL 的执行计划，排查每一步 SQL 执行是否存在问题。所以读懂执行计划是 SQL 优化的先决条件，而了解执行计划的算子是理解 EXPLAIN 命令的关键。
|
||||||
|
|
||||||
|
OceanBase 数据库的执行计划命令有三种模式：EXPLAIN BASIC、EXPLAIN 和 EXPLAIN EXTENDED。这三种模式对执行计划展现不同粒度的细节信息:
|
||||||
|
|
||||||
|
EXPLAIN BASIC 命令用于最基本的计划展示。
|
||||||
|
|
||||||
|
EXPLAIN EXTENDED 命令用于最详细的计划展示（通常在排查问题时使用这种展示模式）。
|
||||||
|
|
||||||
|
EXPLAIN 命令所展示的信息可以帮助普通用户了解整个计划的执行方式。
|
||||||
|
|
||||||
|
EXPLAIN 命令格式如下：
|
||||||
|
EXPLAIN [BASIC | EXTENDED | PARTITIONS | FORMAT = format_name] [PRETTY | PRETTY_COLOR] explainable_stmt
|
||||||
|
format_name:
|
||||||
|
{ TRADITIONAL | JSON }
|
||||||
|
explainable_stmt:
|
||||||
|
{ SELECT statement
|
||||||
|
| DELETE statement
|
||||||
|
| INSERT statement
|
||||||
|
| REPLACE statement
|
||||||
|
| UPDATE statement }
|
||||||
|
|
||||||
|
|
||||||
|
EXPLAIN 命令适用于 SELECT、DELETE、INSERT、REPLACE 和 UPDATE 语句，显示优化器所提供的有关语句执行计划的信息，包括如何处理该语句，如何联接表以及以何种顺序联接表等信息。
|
||||||
|
|
||||||
|
一般来说，可以使用 EXPLAIN EXTENDED 命令，将表扫描的范围段展示出来。使用 EXPLAIN OUTLINE 命令可以显示 Outline 信息。
|
||||||
|
|
||||||
|
FORMAT 选项可用于选择输出格式。TRADITIONAL 表示以表格格式显示输出，这也是默认设置。JSON 表示以 JSON 格式显示信息。
|
||||||
|
|
||||||
|
使用 EXPLAIN PARTITITIONS 也可用于检查涉及分区表的查询。如果检查针对非分区表的查询，则不会产生错误，但 PARTIONS 列的值始终为 NULL。
|
||||||
|
|
||||||
|
对于复杂的执行计划，可以使用 PRETTY 或者 PRETTY_COLOR 选项将计划树中的父节点和子节点使用树线或彩色树线连接起来，使得执行计划展示更方便阅读。示例如下：
|
||||||
|
obclient> CREATE TABLE p1table(c1 INT ,c2 INT) PARTITION BY HASH(c1) PARTITIONS 2;
|
||||||
|
Query OK, 0 rows affected
|
||||||
|
|
||||||
|
obclient> CREATE TABLE p2table(c1 INT ,c2 INT) PARTITION BY HASH(c1) PARTITIONS 4;
|
||||||
|
Query OK, 0 rows affected
|
||||||
|
|
||||||
|
obclient> EXPLAIN EXTENDED PRETTY_COLOR SELECT * FROM p1table p1 JOIN p2table p2 ON p1.c1=p2.c2\G
|
||||||
|
*************************** 1. row ***************************
|
||||||
|
Query Plan: ==========================================================
|
||||||
|
|ID|OPERATOR |NAME |EST. ROWS|COST|
|
||||||
|
----------------------------------------------------------
|
||||||
|
|0 |PX COORDINATOR | |1 |278 |
|
||||||
|
|1 | EXCHANGE OUT DISTR |:EX10001|1 |277 |
|
||||||
|
|2 | HASH JOIN | |1 |276 |
|
||||||
|
|3 | ├PX PARTITION ITERATOR | |1 |92 |
|
||||||
|
|4 | │ TABLE SCAN |P1 |1 |92 |
|
||||||
|
|5 | └EXCHANGE IN DISTR | |1 |184 |
|
||||||
|
|6 | EXCHANGE OUT DISTR (PKEY)|:EX10000|1 |184 |
|
||||||
|
|7 | PX PARTITION ITERATOR | |1 |183 |
|
||||||
|
|8 | TABLE SCAN |P2 |1 |183 |
|
||||||
|
==========================================================
|
||||||
|
|
||||||
|
Outputs & filters:
|
||||||
|
-------------------------------------
|
||||||
|
0 - output([INTERNAL_FUNCTION(P1.C1, P1.C2, P2.C1, P2.C2)]), filter(nil)
|
||||||
|
1 - output([INTERNAL_FUNCTION(P1.C1, P1.C2, P2.C1, P2.C2)]), filter(nil), dop=1
|
||||||
|
2 - output([P1.C1], [P2.C2], [P1.C2], [P2.C1]), filter(nil),
|
||||||
|
equal_conds([P1.C1 = P2.C2]), other_conds(nil)
|
||||||
|
3 - output([P1.C1], [P1.C2]), filter(nil)
|
||||||
|
4 - output([P1.C1], [P1.C2]), filter(nil),
|
||||||
|
access([P1.C1], [P1.C2]), partitions(p[0-1])
|
||||||
|
5 - output([P2.C2], [P2.C1]), filter(nil)
|
||||||
|
6 - (#keys=1, [P2.C2]), output([P2.C2], [P2.C1]), filter(nil), dop=1
|
||||||
|
7 - output([P2.C1], [P2.C2]), filter(nil)
|
||||||
|
8 - output([P2.C1], [P2.C2]), filter(nil),
|
||||||
|
access([P2.C1], [P2.C2]), partitions(p[0-3])
|
||||||
|
|
||||||
|
1 row in set
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
## 执行计划形状与算子信息
|
||||||
|
|
||||||
|
在数据库系统中，执行计划在内部通常是以树的形式来表示的，但是不同的数据库会选择不同的方式展示给用户。
|
||||||
|
|
||||||
|
如下示例分别为 PostgreSQL 数据库、Oracle 数据库和 OceanBase 数据库对于 TPCDS Q3 的计划展示。
|
||||||
|
|
||||||
|
```sql
|
||||||
|
obclient> SELECT /*TPC-DS Q3*/ *
|
||||||
|
FROM (SELECT dt.d_year,
|
||||||
|
item.i_brand_id brand_id,
|
||||||
|
item.i_brand brand,
|
||||||
|
Sum(ss_net_profit) sum_agg
|
||||||
|
FROM date_dim dt,
|
||||||
|
store_sales,
|
||||||
|
item
|
||||||
|
WHERE dt.d_date_sk = store_sales.ss_sold_date_sk
|
||||||
|
AND store_sales.ss_item_sk = item.i_item_sk
|
||||||
|
AND item.i_manufact_id = 914
|
||||||
|
AND dt.d_moy = 11
|
||||||
|
GROUP BY dt.d_year,
|
||||||
|
item.i_brand,
|
||||||
|
item.i_brand_id
|
||||||
|
ORDER BY dt.d_year,
|
||||||
|
sum_agg DESC,
|
||||||
|
brand_id)
|
||||||
|
WHERE ROWNUM <= 100;
|
||||||
|
|
||||||
|
PostgreSQL 数据库执行计划展示如下：
|
||||||
|
Limit (cost=13986.86..13987.20 rows=27 width=91)
|
||||||
|
Sort (cost=13986.86..13986.93 rows=27 width=65)
|
||||||
|
Sort Key: dt.d_year, (sum(store_sales.ss_net_profit)), item.i_brand_id
|
||||||
|
HashAggregate (cost=13985.95..13986.22 rows=27 width=65)
|
||||||
|
Merge Join (cost=13884.21..13983.91 rows=204 width=65)
|
||||||
|
Merge Cond: (dt.d_date_sk = store_sales.ss_sold_date_sk)
|
||||||
|
Index Scan using date_dim_pkey on date_dim dt (cost=0.00..3494.62 rows=6080 width=8)
|
||||||
|
Filter: (d_moy = 11)
|
||||||
|
Sort (cost=12170.87..12177.27 rows=2560 width=65)
|
||||||
|
Sort Key: store_sales.ss_sold_date_sk
|
||||||
|
Nested Loop (cost=6.02..12025.94 rows=2560 width=65)
|
||||||
|
Seq Scan on item (cost=0.00..1455.00 rows=16 width=59)
|
||||||
|
Filter: (i_manufact_id = 914)
|
||||||
|
Bitmap Heap Scan on store_sales (cost=6.02..658.94 rows=174 width=14)
|
||||||
|
Recheck Cond: (ss_item_sk = item.i_item_sk)
|
||||||
|
Bitmap Index Scan on store_sales_pkey (cost=0.00..5.97 rows=174 width=0)
|
||||||
|
Index Cond: (ss_item_sk = item.i_item_sk)
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
Oracle 数据库执行计划展示如下：
|
||||||
|
Plan hash value: 2331821367
|
||||||
|
--------------------------------------------------------------------------------------------------
|
||||||
|
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
|
||||||
|
--------------------------------------------------------------------------------------------------
|
||||||
|
| 0 | SELECT STATEMENT | | 100 | 9100 | 3688 (1)| 00:00:01 |
|
||||||
|
|* 1 | COUNT STOPKEY | | | | | |
|
||||||
|
| 2 | VIEW | | 2736 | 243K| 3688 (1)| 00:00:01 |
|
||||||
|
|* 3 | SORT ORDER BY STOPKEY | | 2736 | 256K| 3688 (1)| 00:00:01 |
|
||||||
|
| 4 | HASH GROUP BY | | 2736 | 256K| 3688 (1)| 00:00:01 |
|
||||||
|
|* 5 | HASH JOIN | | 2736 | 256K| 3686 (1)| 00:00:01 |
|
||||||
|
|* 6 | TABLE ACCESS FULL | DATE_DIM | 6087 | 79131 | 376 (1)| 00:00:01 |
|
||||||
|
| 7 | NESTED LOOPS | | 2865 | 232K| 3310 (1)| 00:00:01 |
|
||||||
|
| 8 | NESTED LOOPS | | 2865 | 232K| 3310 (1)| 00:00:01 |
|
||||||
|
|* 9 | TABLE ACCESS FULL | ITEM | 18 | 1188 | 375 (0)| 00:00:01 |
|
||||||
|
|* 10 | INDEX RANGE SCAN | SYS_C0010069 | 159 | | 2 (0)| 00:00:01 |
|
||||||
|
| 11 | TABLE ACCESS BY INDEX ROWID| STORE_SALES | 159 | 2703 | 163 (0)| 00:00:01 |
|
||||||
|
--------------------------------------------------------------------------------------------------
|
||||||
|
|
||||||
|
OceanBase 数据库执行计划展示如下：
|
||||||
|
|ID|OPERATOR |NAME |EST. ROWS|COST |
|
||||||
|
-------------------------------------------------------
|
||||||
|
|0 |LIMIT | |100 |81141|
|
||||||
|
|1 | TOP-N SORT | |100 |81127|
|
||||||
|
|2 | HASH GROUP BY | |2924 |68551|
|
||||||
|
|3 | HASH JOIN | |2924 |65004|
|
||||||
|
|4 | SUBPLAN SCAN |VIEW1 |2953 |19070|
|
||||||
|
|5 | HASH GROUP BY | |2953 |18662|
|
||||||
|
|6 | NESTED-LOOP JOIN| |2953 |15080|
|
||||||
|
|7 | TABLE SCAN |ITEM |19 |11841|
|
||||||
|
|8 | TABLE SCAN |STORE_SALES|161 |73 |
|
||||||
|
|9 | TABLE SCAN |DT |6088 |29401|
|
||||||
|
=======================================================
|
||||||
|
|
||||||
|
由示例可见，OceanBase 数据库的计划展示与 Oracle 数据库类似。
|
||||||
|
|
||||||
|
OceanBase 数据库执行计划中的各列的含义如下：
|
||||||
|
列名 含义
|
||||||
|
ID 执行树按照前序遍历的方式得到的编号（从 0 开始）。
|
||||||
|
OPERATOR 操作算子的名称。
|
||||||
|
NAME 对应表操作的表名（索引名）。
|
||||||
|
EST. ROWS 估算该操作算子的输出行数。
|
||||||
|
COST 该操作算子的执行代价（微秒）。
|
||||||
|
|
||||||
|
|
||||||
|
OceanBase 数据库 EXPLAIN 命令输出的第一部分是执行计划的树形结构展示。其中每一个操作在树中的层次通过其在 operator 中的缩进予以展示，层次最深的优先执行，层次相同的以特定算子的执行顺序为标准来执行。
|
||||||
|
|
||||||
|
问题: update a not exists (b…)
|
||||||
|
我一开始以为 B是驱动表，B的数据挺多的 后来看到NLAJ，是说左边的表关联右边的表
|
||||||
|
所以这个的驱动表是不是实际是A，用A的匹配B的，这个理解有问题吗
|
||||||
|
|
||||||
|
回答: 没错 A 驱动 B的
|
||||||
|
|
||||||
|
问题: 光知道最下最右的是驱动表了 所以一开始搞得有点懵 :sweat_smile:
|
||||||
|
|
||||||
|
回答: nlj应该原理应该都是左表(驱动表)的记录探测右表(被驱动表)， 选哪张成为左表或右表就基于一些其他考量了，比如数据量， 而anti join/semi join只是对 not exist/exist的一种优化，相关的原理和资料网上可以查阅一下
|
||||||
|
|
||||||
|
问题: 也就是nlj 就是按照之前理解的谁先执行 谁就是驱动表 也就是执行计划中的最右的表
|
||||||
|
而anti join/semi join，谁在not exist左面，谁就是驱动表。这么理解对吧
|
||||||
|
|
||||||
|
回答: nlj也是左表的表是驱动表，这个要了解下计划执行方面的基本原理，取左表的一行数据，再遍历右表，一旦满足连接条件，就可以返回数据
|
||||||
|
anti/semi只是因为not exists/exist的语义只是返回左表数据，改成anti join是一种计划优化，连接的方式比子查询更优
|
||||||
|
"""
|
||||||
text_list = [base_knowledge]
|
text_list = [base_knowledge]
|
||||||
index = knowledged_qa_demo(text_list)
|
index = knowledged_qa_demo(text_list)
|
||||||
response = index.query(q)
|
response = index.query(q)
|
||||||
|
|||||||
@@ -4,10 +4,10 @@
|
|||||||
import torch
|
import torch
|
||||||
|
|
||||||
@torch.inference_mode()
|
@torch.inference_mode()
|
||||||
def generate_output(model, tokenizer, params, device, context_len=2048):
|
def generate_output(model, tokenizer, params, device, context_len=2048, stream_interval=2):
|
||||||
prompt = params["prompt"]
|
prompt = params["prompt"]
|
||||||
temperature = float(params.get("temperature", 1.0))
|
temperature = float(params.get("temperature", 1.0))
|
||||||
max_new_tokens = int(params.get("max_new_tokens", 256))
|
max_new_tokens = int(params.get("max_new_tokens", 1024))
|
||||||
stop_parameter = params.get("stop", None)
|
stop_parameter = params.get("stop", None)
|
||||||
if stop_parameter == tokenizer.eos_token:
|
if stop_parameter == tokenizer.eos_token:
|
||||||
stop_parameter = None
|
stop_parameter = None
|
||||||
@@ -21,29 +21,29 @@ def generate_output(model, tokenizer, params, device, context_len=2048):
|
|||||||
else:
|
else:
|
||||||
raise TypeError("Stop parameter must be string or list of strings.")
|
raise TypeError("Stop parameter must be string or list of strings.")
|
||||||
|
|
||||||
pos = -1
|
|
||||||
input_ids = tokenizer(prompt).input_ids
|
input_ids = tokenizer(prompt).input_ids
|
||||||
output_ids = []
|
output_ids = []
|
||||||
|
|
||||||
max_src_len = context_len - max_new_tokens - 8
|
max_src_len = context_len - max_new_tokens - 8
|
||||||
input_ids = input_ids[-max_src_len:]
|
input_ids = input_ids[-max_src_len:]
|
||||||
|
stop_word = None
|
||||||
|
|
||||||
for i in range(max_new_tokens):
|
for i in range(max_new_tokens):
|
||||||
if i == 0:
|
if i == 0:
|
||||||
out = model(torch.as_tensor([input_ids], device=device), use_cache=True)
|
out = model(
|
||||||
|
torch.as_tensor([input_ids], device=device), use_cache=True)
|
||||||
logits = out.logits
|
logits = out.logits
|
||||||
past_key_values = out.past_key_values
|
past_key_values = out.past_key_values
|
||||||
else:
|
else:
|
||||||
out = model(
|
out = model(input_ids=torch.as_tensor([[token]], device=device),
|
||||||
input_ids=torch.as_tensor([[token]], device=device),
|
use_cache=True,
|
||||||
use_cache=True,
|
past_key_values=past_key_values)
|
||||||
past_key_values=past_key_values,
|
|
||||||
)
|
|
||||||
logits = out.logits
|
logits = out.logits
|
||||||
past_key_values = out.past_key_values
|
past_key_values = out.past_key_values
|
||||||
|
|
||||||
last_token_logits = logits[0][-1]
|
last_token_logits = logits[0][-1]
|
||||||
|
|
||||||
|
|
||||||
if temperature < 1e-4:
|
if temperature < 1e-4:
|
||||||
token = int(torch.argmax(last_token_logits))
|
token = int(torch.argmax(last_token_logits))
|
||||||
else:
|
else:
|
||||||
@@ -57,15 +57,22 @@ def generate_output(model, tokenizer, params, device, context_len=2048):
|
|||||||
else:
|
else:
|
||||||
stopped = False
|
stopped = False
|
||||||
|
|
||||||
|
|
||||||
output = tokenizer.decode(output_ids, skip_special_tokens=True)
|
output = tokenizer.decode(output_ids, skip_special_tokens=True)
|
||||||
|
# print("Partial output:", output)
|
||||||
for stop_str in stop_strings:
|
for stop_str in stop_strings:
|
||||||
|
# print(f"Looking for '{stop_str}' in '{output[:l_prompt]}'#END")
|
||||||
pos = output.rfind(stop_str)
|
pos = output.rfind(stop_str)
|
||||||
if pos != -1:
|
if pos != -1:
|
||||||
|
# print("Found stop str: ", output)
|
||||||
output = output[:pos]
|
output = output[:pos]
|
||||||
|
# print("Trimmed output: ", output)
|
||||||
stopped = True
|
stopped = True
|
||||||
|
stop_word = stop_str
|
||||||
break
|
break
|
||||||
else:
|
else:
|
||||||
pass
|
pass
|
||||||
|
# print("Not found")
|
||||||
|
|
||||||
if stopped:
|
if stopped:
|
||||||
break
|
break
|
||||||
|
|||||||
@@ -17,17 +17,25 @@ class VicunaRequestLLM(LLM):
|
|||||||
if isinstance(stop, list):
|
if isinstance(stop, list):
|
||||||
stop = stop + ["Observation:"]
|
stop = stop + ["Observation:"]
|
||||||
|
|
||||||
|
skip_echo_len = len(prompt.replace("</s>", " ")) + 1
|
||||||
params = {
|
params = {
|
||||||
"prompt": prompt,
|
"prompt": prompt,
|
||||||
"temperature": 0,
|
"temperature": 0.7,
|
||||||
"max_new_tokens": 256,
|
"max_new_tokens": 1024,
|
||||||
"stop": stop
|
"stop": stop
|
||||||
}
|
}
|
||||||
response = requests.post(
|
response = requests.post(
|
||||||
url=urljoin(vicuna_model_server, self.vicuna_generate_path),
|
url=urljoin(vicuna_model_server, self.vicuna_generate_path),
|
||||||
data=json.dumps(params)
|
data=json.dumps(params),
|
||||||
)
|
)
|
||||||
response.raise_for_status()
|
response.raise_for_status()
|
||||||
|
# for chunk in response.iter_lines(decode_unicode=False, delimiter=b"\0"):
|
||||||
|
# if chunk:
|
||||||
|
# data = json.loads(chunk.decode())
|
||||||
|
# if data["error_code"] == 0:
|
||||||
|
# output = data["text"][skip_echo_len:].strip()
|
||||||
|
# output = self.post_process_code(output)
|
||||||
|
# yield output
|
||||||
return response.json()["response"]
|
return response.json()["response"]
|
||||||
|
|
||||||
@property
|
@property
|
||||||
|
|||||||
@@ -115,6 +115,5 @@ def embeddings(prompt_request: EmbeddingRequest):
|
|||||||
return {"response": [float(x) for x in output]}
|
return {"response": [float(x) for x in output]}
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
if __name__ == "__main__":
|
if __name__ == "__main__":
|
||||||
uvicorn.run(app, host="0.0.0.0", log_level="info")
|
uvicorn.run(app, host="0.0.0.0", log_level="info")
|
||||||
@@ -124,6 +124,7 @@ def http_bot(state, temperature, max_new_tokens, request: gr.Request):
|
|||||||
start_tstamp = time.time()
|
start_tstamp = time.time()
|
||||||
model_name = LLM_MODEL
|
model_name = LLM_MODEL
|
||||||
|
|
||||||
|
# TODO 这里的请求需要拼接现有知识库, 使得其根据现有知识库作答, 所以prompt需要继续优化
|
||||||
if state.skip_next:
|
if state.skip_next:
|
||||||
# This generate call is skipped due to invalid inputs
|
# This generate call is skipped due to invalid inputs
|
||||||
yield (state, state.to_gradio_chatbot()) + (no_change_btn,) * 5
|
yield (state, state.to_gradio_chatbot()) + (no_change_btn,) * 5
|
||||||
|
|||||||
Reference in New Issue
Block a user