Empathy is widely regarded as an inherently positive feature of supportive online interactions, but its value is shaped by context. In this study, we argue that empathy should be understood not as a uniform good but as a multidimensional and relational practice. Rather than treating empathy as a binary or unidimensional attribute, we propose a new framework that captures how empathy is solicited in posts and how it is expressed in replies, emphasising that context is critical in determining its appropriateness and effectiveness. Using post-reply data from six communities across Reddit and Stack Exchange, we conduct a three-phase study. First, we develop a fine-grained annotation framework grounded in psychological theory to capture distinct empathy practices in both opening posts and replies. Second, we fine-tune language models to detect these practices, showing that they can effectively capture the distinct empathy practices in our schema. Third, we apply the models at scale and conduct an empirical analysis to examine platform- and community-specific patterns in how empathy is both elicited and expressed. Our findings challenge current assumptions about online empathy and offer a more contextualised understanding of its role in online discourse. We identify future directions for platform design and contextual support for community members.