1、环境
Spring Data JPA、Spring Boot
2、利用HQL实现动态条件、分页查询。
(1)DAO中的定义
/** * 分页查询预警信息(注:三个对象并无建立对象引用关系,即对应表之间并没有建立外键关联,只是存储了ID值) * @param yjfl 预警分类 * @param yjjb 预警级别 * @param bq 标签 * @param beginDate 预警时间(开始时间) * @param endDate 预警时间(结束时间) * @param xm 人员姓名 * @param gmsfzhm 公民身份证号码 * @param pid 人员ID * @param pageable 分页信息 * @return */ @Query( value = "select t,p from WarningInfo t, BigRelation b, Person p" + " where t.id = b.ktId" + " and b.ztId = p.id" + " and (t.yjfl = ?1 or ?1 is null)" + " and (t.yjjb = ?2 or ?2 is null)" + " and (t.bq like ?3 or ?3 is null)" + " and (t.sj >= ?4 or ?4 is null)" + " and (t.sj <= ?5 or ?5 is null)" + " and b.gx = 'GX_R_YJXX_ZT'" + " and (p.xm like ?6 or ?6 is null)" + " and (p.gmsfzhm = ?7 or ?7 is null)" + " and (p.id = ?8 or ?8 is null)" + " order by ?#{#pageable}", countQuery = "select count(*) from WarningInfo t, BigRelation b, Person p" + " where t.id = b.ktId" + " and b.ztId = p.id" + " and (t.yjfl = ?1 or ?1 is null)" + " and (t.yjjb = ?2 or ?2 is null)" + " and (t.bq like ?3 or ?3 is null)" + " and (t.sj >= ?4 or ?4 is null)" + " and (t.sj <= ?5 or ?5 is null)" + " and b.gx = 'GX_R_YJXX_ZT'" + " and (p.xm like ?6 or ?6 is null)" + " and (p.gmsfzhm = ?7 or ?7 is null)" + " and (p.id = ?8 or ?8 is null)" + " order by ?#{#pageable}" ) Page<Object[]> findWarningInfo(String yjfl, String yjjb, String bq, Date beginDate,
Date endDate, String xm, String gmsfzhm, Long pid, Pageable pageable);
(2)调用示例
public Result queryYj(final WarningInfoVo vo) { Sort sort = new Sort("desc".equals(vo.getOrder())?Direction.DESC:Direction.ASC, vo.getSort()); // 排序 Pageable pageable = new PageRequest(vo.getPage() - 1, vo.getRows(), sort); // 分页查询条件 String yjfl = vo.getYjfl(); //预警分类 String yjjb = vo.getYjjb(); //预警级别 String bq = vo.getBq(); //标签 String xm = vo.getXm(); //人员姓名 String gmsfzhm = vo.getGmsfzhm(); //公民身份证号码 String sj_start = vo.getSj_start(); //开始时间 String sj_end = vo.getSj_end(); //结束时间 //条件加工 if(!xx.isEmpty(bq)) { bq = "%," + bq + ",%"; } if(!xx.isEmpty(xm)) { xm = "%" + xm + "%"; } Date beginDate = null; if(!xx.isEmpty(sj_start)) { beginDate = xx.toDate(sj_start); } Date endDate = null; if(!xx.isEmpty(sj_end)) { endDate = xx.toTime(sj_end + " 23:59:59"); } // 查询 Page<Object[]> page = dao.findWarningInfo(yjfl, yjjb, bq, beginDate, endDate, xm, gmsfzhm, null, pageable); //转换构建列表数据 List<Object[]> arrList = page.getContent(); List<WarningInfoDto> list = new ArrayList<>(arrList.size()); for(Object[] arr: arrList) { list.add(new WarningInfoDto((WarningInfo)arr[0], (Person)arr[1])); } //构建返回结果 Result result =new Result(); result.setRows(list); result.setTotal(page.getTotalElements()); return result; }
3、分组查询
(1)DAO中的定义
@Query("select count(*) as num, t.hy as hy from DataResource t where t.state.code='06'group by t.hy order by t.hy.orderNum ") List<Object> findGroupByHy(); @Query("select count(*) as num, t.yw as yw from DataResource t where t.state.code='06' group by t.yw order by t.yw.orderNum") List<Object> findGroupByYw();
(2)调用示例
public List<StatisticsVo> statisticsThisLv1Group(boolean isYw)
{ List<StatisticsVo> list = new ArrayList<>(); if (isYw) {//业务 List<Object> _list = dao.findGroupByYw(); for(Object row:_list) { Object[] cells = (Object[]) row; Long num = (Long) cells[0]; CodeBusinessType sort = (CodeBusinessType) cells[1]; System.out.println(sort.getName()+" "+num); StatisticsVo vo = new StatisticsVo(); vo.setCount(num); vo.setCode(sort.getId() + ""); vo.setName(sort.getName()); list.add(vo); } } else {//行业 List<Object> _list = dao.findGroupByHy(); for(Object row:_list) { Object[] cells = (Object[]) row; Long num = (Long) cells[0]; CodeIndustry sort = (CodeIndustry) cells[1]; System.out.println(sort.getName()+" "+num); StatisticsVo vo = new StatisticsVo(); vo.setCount(num); vo.setCode(sort.getId() + ""); vo.setName(sort.getName()); list.add(vo); } } return list; }