Log-based anomaly detection plays a crucial role in ensuring the stability of software. However, current approaches for log-based anomaly detection heavily depend on a vast amount of labeled historical data, which is often unavailable in many real-world systems. To mitigate this problem, we leverage the features of the abundant historical labeled logs of mature systems to help construct anomaly detection models of new systems with very few labels, that is, to generalize the model ability trained from labeled logs of mature systems to achieve anomaly detection on new systems with insufficient data labels. Specifically, we propose MetaLog, a generalizable cross-system anomaly detection approach. MetaLog first incorporates a globally consistent semantic embedding module to obtain log event semantic embedding vectors in a shared global space. Then it leverages the meta-learning paradigm to improve the model's generalization ability. We evaluate MetaLog's performance on four public log datasets (HDFS, BGL, OpenStack, and Thunderbird) from four different systems. Results show that MetaLog reaches over 80% F1-score when using only 1% labeled logs of the target system, showing similar performance with state-of-the-art supervised anomaly detection models trained with 100% labeled data. Besides, it outperforms state-of-art transfer-learning-based cross-system anomaly detection models by 20% in the same settings of 1% labeled training logs of the target system.