Optimize go N steps over edge #1471
Conversation
dangleptr
requested review from
monadbobo,
dutor,
darionyaphet,
sherman-the-tank,
CPWstatic,
critical27,
laura-ding,
whitewum,
liuyu85cn and
bright-starry-sky
|
Unit testing passed. |
|
Unit testing passed. |
|
Unit testing passed. |
|
Unit testing failed. |
|
Unit testing passed. |
|
Unit testing passed. |
1 similar comment
|
Unit testing passed. |
| int64_t totalRows = 0; | ||
| for (auto& resp : rpcResp.responses()) { | ||
| if (resp.get_total_edges() != nullptr) { | ||
| totalRows += *resp.get_total_edges(); |
There was a problem hiding this comment.
maybe *(resp.get_total_edges()) more readable
src/graph/GoExecutor.cpp
Outdated
| if (!edgeSchema.empty()) { | ||
| auto it = edgeSchema.find(edgeType); | ||
| DCHECK(it != edgeSchema.end()); | ||
| reader = RowReader::getRowReader(edge.props, it->second); |
There was a problem hiding this comment.
because the reader depend edgeSchema, if the edgeSchema is empty, the reader is nullptr, then CHECK(reader != nullptr); will crash , so if can we return directly when edgeSchema is empty?
There was a problem hiding this comment.
If getAliasProp invoked, it means the reader should not be NULL. Otherwise it is a bug.
|
close #1604 |
| @@ -1408,5 +1468,73 @@ SupportedType GoExecutor::getPropTypeFromInterim(const std::string &prop) const | |||
| return index_->getColumnType(prop); | |||
| } | |||
|
|
|||
| nebula::cpp2::SupportedType GoExecutor::calculateExprType(Expression* exp) const { | |||
There was a problem hiding this comment.
If the schema cache of graphd and the storaged are different, maybe has problem.
There was a problem hiding this comment.
Yes, we use the schema in graphd.
And the problem also existed in current master branch, because different storaged maybe have different versions schema when updating it.
| << time::WallClock::fastNowInMicroSec() - start << "us"; | ||
| } | ||
| if (!ret.ok()) { | ||
| LOG(ERROR) << "Get rows failed: " << ret.status(); |
Optimize go executor. The pr improve the perf for "go 2 steps over edge" about 20x with about 1M vertex ids returned. Of course, it will improve the perf for other queries, but i have't test it yet. Main changes: 1. Modify the structure storaged returned, separate vertexId from props to avoid "getPropByName" called when getting dstId. (This is a normal request for go) 2. Return directly if current executor is the rightest. It avoid one extra encode/decode procedure which cost lots of time if we have many rows returned. 3. Pre calculate the yield columns types to avoid calculate them for every edge.
Optimize go executor.
The pr improve the perf for "go 2 steps over edge" about 20x with about 1M vertex ids returned.
Of course, it will improve the perf for other queries, but i have't test it yet.
Main changes:
Modify the structure storaged returned, separate vertexId from props to avoid "getPropByName" called when getting dstId. (This is a normal request for go)
Return directly if current executor is the rightest. It avoid one extra encode/decode procedure which cost lots of time if we have many rows returned.
Pre calculate the yield columns types to avoid calculate them for every edge.